從去年開始，出於個人興趣，我們開啟了專案「OpenLandLord」，並構建了规则引擎、数据生成、蒸馏管线、对局评测与可视化后端等模塊。在基礎模塊搭建好後，我們開展了對「LandlordFormer-Lite」的設計。這是一個以 DouZero-ADP 為參考，規模更小、適合在 Apple Silicon 上迭代與部署的鬥地主模型。

截至目前，我們所構建的 LandlordFormer-Lite 只有 139,021 個參數，約為三個 DouZero 角色模型總參數的 3.10%。儘管在實際對局能力中其還沒有追平教師，但已經使我們明確後續改進方向。在本階段工作中，我們通過分析研究後，將價值頭從「只預測輸贏」修改為「預測經截斷的終局分差」，從而使得產物在三組全新牌局上的勝率由 34.26% 提升至 35.19%，平均每局分數由 -2.406 改善至 -2.321。

這是目前最好的階段性結果，但它與我們設想的完整版本仍有距離：三個角色尚未同時受益、叫牌還沒有特化訓練，牌力也需要繼續提高。

為什麼要重新做一個小模型

從理論上來分析，鬥地主對小模型並不友善，主要有四個原因：

1. 資訊不完整。 玩家只能看到自己的手牌與公開歷史，對手手牌始終是隱變數。
2. 角色不對稱。 地主獨自對抗兩名農民，三個座位的決策分布差異很大。
3. 合法動作數量變化很大。 某些局面幾乎是強制出牌，某些地主局面卻可能同時有十幾種選擇。
4. 輸贏不是全部。 炸彈、春天與倍數會放大終局分差；同樣是輸，代價可能完全不同。

DouZero 已經證明，對所有合法動作逐一評分是處理這類問題的有效方式 [@zha2021douzero]。但原始系統為三個角色保留三套模型，重點也在大規模自我對弈。我們想探索另一條路：讓三個角色共享大部分表示，用一個很小的 Transformer 同時理解公開狀態、合法動作集合與角色差異。

目標並不是單純追求最少參數。真正希望得到的是一個可以快速迭代的研究載體：模型足夠小，能反覆做消融；推論足夠快，可以接進即時對局；同時仍保留足夠牌力，讓架構與訓練方法的差異能在完整牌局中顯現。

我們是怎麼做的

1. 讓教師評分所有合法動作

如果蒸餾資料只記錄 DouZero 最後選了哪一手牌，學生只能看到一個硬答案，其他合法動作之間的差異都會消失；保留教師的軟分布也是策略蒸餾的核心方法之一 [@rusu2016policy]。因此，每個決策狀態都保存：

• 當前玩家能看到的狀態與手牌；
• 完整的標準合法動作集合；
• DouZero 對每個合法動作的分數與溫度化機率；
• 教師最後選擇的動作；
• 這副牌的終局結果與分差。

在這套分工中，OpenLandLord 是唯一的規則裁判；LandlordFormer-Lite 學習如何在合法動作中做決策。DouZero-ADP 負責為裁判給出的合法動作提供教師分數，不會取得對手私有手牌，也不能繞過規則引擎。資料則按整副牌切分，避免同一副牌的不同決策同時出現在訓練與驗證中。

目前主要資料由 500 手教師軌跡與 500 手學生／教師混合軌跡構成，共 35,672 個決策狀態。在混合軌跡中，學生控制一半出牌，DouZero 仍會為所有到訪狀態重新提供完整動作分數。這讓資料同時包含教師熟悉的狀態，以及學生自己走偏後真正會遇到的狀態。

2. 用 Transformer 編碼狀態，用 Deep Sets 理解合法動作

LandlordFormer-Lite 的狀態編碼器只有兩層 Transformer：

• hidden size 64；
• 4 個注意力頭；
• feed-forward size 128；
• 15 個手牌牌階／張數 token；
• 三位玩家的公開已出牌摘要；
• 叫牌、出牌與 pass 的公開歷史。

每個合法動作會獨立編碼，並以 query 的方式讀取狀態序列。接著，模型用 masked mean/max Deep Sets 摘要整個合法動作集合，讓策略知道「當前這個動作與其他選項相比如何」，又不依賴合法動作的排列順序；這一設計也延續了集合模型對排列不變性的處理思路 [@lee2019settransformer]。

我們也測過誘導注意力、三套角色專屬策略頭與額外動作分數回歸頭。它們都增加了模型複雜度，但沒有換來更好的完整對局結果。最後保留下來的是秩為 8 的低秩角色 adapter：大部分知識共享，只用很小的殘差表示地主、地主下家與地主上家的差異。

這個組合把模型控制在 139,021 個參數：

這裡得到的第一個方向很明確：對小模型而言，讓架構符合問題本身，比直接增加 Transformer 容量更有效。 合法動作是集合，角色不同但共享大量知識；把這兩點直接寫進模型，比多加一層通用模組更划算。

3. 用牌數守恆限制公開信念

只看公開歷史仍不足以描述不完全資訊。受到 PerfectDou、DouZero+、BAD 與 ReBeL 的啟發 [@yang2022perfectdou; @zhao2022douzeroplus; @foerster2019bad; @brown2020rebel]，我們加入一個很小的公開信念頭。

對每個牌階 $r$，自己的手牌與已出牌可以確定尚未觀察到的牌數 $u_r$。模型預測下一位對手持有 $k_r$ 張的分布：

$$ P(k_r \mid h_{\mathrm{public}}), \qquad 0 \le k_r \le u_r. $$

超過 $u_r$ 的張數直接遮罩；另一位對手的期望張數以 $u_r-E[k_r]$ 計算，因此每個牌階都嚴格符合牌數守恆。策略只接收兩位對手的期望張數，不接收完整類別分布，而且信念特徵在輸入策略前停止梯度。

這幾個限制看似保守，結果卻比更豐富的版本好。只把信念當輔助任務、把完整類別分布輸入策略，或再加入二階矩，都沒有穩定改善牌力。最有效的形式反而是「守恆、期望值、停止梯度」。

這指向第二個方向：不完全資訊模型需要的未必是更大的信念表示，而可能是更可靠的資訊瓶頸。 下一步值得研究的不是無限制增加隱藏資訊維度，而是如何讓少量信念特徵更準確、更符合規則，也更能服務不同角色。

4. 從「輸或贏」改成「輸贏多少」

前幾個版本用 $y\in{-1,1}$ 監督價值頭。這會把小輸與高倍數慘敗視為同一個目標，無法向表示層傳遞風險差異。

目前最有效的修改，是改用經縮放與截斷的終局分差：

$$ v^* = \operatorname{clip}\left(\frac{\Delta_{\mathrm{terminal}}}{24},-1,1\right). $$

截斷保留了「輸得多重」的訊號，又避免少數極端倍數完全支配梯度。這個修改不增加模型參數，也不改變推論圖，只改變訓練時價值頭看到的目標。

做完之後，結果發生了什麼

探索階段：分數目標帶來最明顯的提升

所有版本都與 DouZero-ADP 使用同一規則引擎與叫牌流程。每組 300 手原始牌會把學生輪流放到三個座位，形成 900 場鏡像對局。統計時仍以原始牌局為單位，避免把共享同一副牌的三場對局視為完全獨立樣本。

前三組牌局用於比較架構與訓練方法，共 2,700 場對局：

分歧重播提高了部分狀態上的教師模仿率，但平均分數反而惡化，並出現幾場極端失分。分數感知價值目標則同時改善勝率與分數，成為最值得繼續測試的版本。

這裡揭示了第三個方向：選模不能只看教師一致率，也不能只看勝率。 完整牌局的分數、角色切片與尾部損失，會暴露局部模仿指標看不到的風險。

全新牌局：方向延續，但角色仍不均衡

選定分數感知版本後，我們再使用三組全新牌局，共 2,700 場鏡像對局：

三組新牌局的勝率與分數全部朝同一方向改善。以原始牌局做配對 bootstrap，勝率差為 +0.93 個百分點，95% 信賴區間為 [-0.26, +2.15]；平均分數差為 +0.084，區間為 [-0.068, +0.239]。這是目前的最優結果，也說明下一輪需要更多全新牌局來縮小波動範圍。

按角色拆開後，新的研究問題也出現了：

地主與地主下家受益，地主上家卻退化。未來的角色共享研究需要在「共享表示」與「角色專屬調整」之間找到更好的位置，而不是簡單拆回三套完整模型。

更大的資料集沒有自然帶來更強的模型

我們也嘗試把資料從 500 手擴大到 1,000 或 2,000 手。在累計更新量相近時，更多不同狀態意味著每個狀態得到的重複優化更少，實戰表現反而下降。1,000 手版本的離線教師一致率更高，實戰卻只有 26.44% 勝率與 -3.157 平均分數。

同樣地，用學生軌跡完全取代教師軌跡會把勝率降到 29.33%；合併兩種來源才回到 31.81%。這表示資料覆蓋與訓練曝光必須分開控制。下一步不能只問「再收集多少手牌」，而要研究「哪些狀態值得更多更新，以及如何在教師狀態和學生訪問狀態之間分配預算」。

小模型已經足以支援即時對局

固定形狀 Core ML 匯出在這次 Apple Silicon 測試中達到 0.466 ms P50、0.623 ms P95：

三種設定的差異很小，CPU_ONLY 反而略快。compute-unit 描述的是調度偏好；若要進一步利用 Neural Engine，還需要逐算子 profiling 與更針對固定形狀算子的設計。對目前的可視化與人機對局而言，這個延遲已經足以讓模型即時回應。

這些結果把下一步指向哪裡

後續研究方向規劃如下。

1. 讓分數目標真正對應高倍數風險

目前有效的尺度設定是 24，下一步會比較不同縮放尺度、未截斷分數與更穩健的分位數目標。重點不只看平均分數，也要單獨追蹤炸彈與春天造成的尾部損失。

2. 改善角色共享，而不是回到三套模型

地主上家退化說明目前的低秩 adapter 還不夠精細。可以研究依角色與合法動作數動態調節的 adapter，或只在高分支地主狀態增加少量容量，同時保留大部分共享參數。

3. 把公開信念變得更準，而不是更大

完整分布與二階矩沒有自然帶來提升。更值得做的是校準期望牌數、利用玩家出牌行為更新信念，並研究信念誤差在三個角色上的不同影響。

4. 同時控制資料覆蓋與更新次數

更多牌局不等於更多有效學習。後續資料策略需要按整副牌控制曝光，針對學生常到訪、教師分歧大且終局代價高的狀態分配額外更新，而不是單純擴大資料集。

5. 把叫牌納入同一套學習問題

現在的結果主要反映出牌策略。完整版本還需要學習叫地主、搶地主與加倍，讓角色形成、倍數風險與後續出牌共享同一套價值尺度。

階段性結語

作為 OpenLandLord 目前的輕量策略模型，LandlordFormer-Lite 用約 3.10% 的教師參數量，實現了快速迭代、Core ML 匯出與完整三座位對局。最有效的改動並不是擴大網路，而是加入符合問題結構的限制：合法動作集合、低秩角色差異、守恆感知公開信念，以及能區分輸贏幅度的價值目標。

目前最優版本在全新牌局上同時改善了勝率與平均分數，但與最初設想相比，仍需要提高整體牌力、解決地主上家的退化，並把叫牌規則整合進來。這些差距使得後續研究從架構變化這個龐大的問題收斂為幾個可以清楚驗證的小問題。

@inproceedings{zha2021douzero,
  author = {Zha, Daochen and Xie, Jingru and Ma, Wenye and Zhang, Sheng and Lian, Xiangru and Hu, Xia and Liu, Ji},
  title = {DouZero: Mastering DouDizhu with Self-Play Deep Reinforcement Learning},
  booktitle = {Proceedings of the 38th International Conference on Machine Learning},
  series = {Proceedings of Machine Learning Research},
  volume = {139},
  pages = {12333--12344},
  publisher = {PMLR},
  year = {2021},
  url = {https://proceedings.mlr.press/v139/zha21a.html}
}

@inproceedings{yang2022perfectdou,
  author = {Yang, Guan and Liu, Minghuan and Hong, Weijun and Zhang, Weinan and Fang, Fei and Zeng, Guangjun and Lin, Yue},
  title = {PerfectDou: Dominating DouDizhu with Perfect Information Distillation},
  booktitle = {Advances in Neural Information Processing Systems},
  volume = {35},
  pages = {34954--34965},
  publisher = {Curran Associates, Inc.},
  year = {2022},
  doi = {10.52202/068431-2533}
}

@inproceedings{zhao2022douzeroplus,
  author = {Zhao, Youpeng and Zhao, Jian and Hu, Xunhan and Zhou, Wengang and Li, Houqiang},
  title = {DouZero+: Improving DouDizhu AI by Opponent Modeling and Coach-guided Learning},
  booktitle = {2022 IEEE Conference on Games},
  pages = {127--134},
  publisher = {IEEE},
  year = {2022},
  doi = {10.1109/CoG51982.2022.9893710}
}

@inproceedings{lee2019settransformer,
  author = {Lee, Juho and Lee, Yoonho and Kim, Jungtaek and Kosiorek, Adam and Choi, Seungjin and Teh, Yee Whye},
  title = {Set Transformer: A Framework for Attention-based Permutation-Invariant Neural Networks},
  booktitle = {Proceedings of the 36th International Conference on Machine Learning},
  series = {Proceedings of Machine Learning Research},
  volume = {97},
  pages = {3744--3753},
  publisher = {PMLR},
  year = {2019},
  url = {https://proceedings.mlr.press/v97/lee19d.html}
}

@inproceedings{rusu2016policy,
  author = {Rusu, Andrei A. and Colmenarejo, Sergio Gomez and Gülçehre, Çaglar and Desjardins, Guillaume and Kirkpatrick, James and Pascanu, Razvan and Mnih, Volodymyr and Kavukcuoglu, Koray and Hadsell, Raia},
  title = {Policy Distillation},
  booktitle = {International Conference on Learning Representations},
  year = {2016},
  url = {https://mlanthology.org/iclr/2016/rusu2016iclr-policy/}
}

@inproceedings{foerster2019bad,
  author = {Foerster, Jakob and Song, Francis and Hughes, Edward and Burch, Neil and Dunning, Iain and Whiteson, Shimon and Botvinick, Matthew and Bowling, Michael},
  title = {Bayesian Action Decoder for Deep Multi-Agent Reinforcement Learning},
  booktitle = {Proceedings of the 36th International Conference on Machine Learning},
  series = {Proceedings of Machine Learning Research},
  volume = {97},
  pages = {1942--1951},
  publisher = {PMLR},
  year = {2019},
  url = {https://proceedings.mlr.press/v97/foerster19a.html}
}

@inproceedings{brown2020rebel,
  author = {Brown, Noam and Bakhtin, Anton and Lerer, Adam and Gong, Qucheng},
  title = {Combining Deep Reinforcement Learning and Search for Imperfect-Information Games},
  booktitle = {Advances in Neural Information Processing Systems},
  volume = {33},
  pages = {17057--17069},
  publisher = {Curran Associates, Inc.},
  year = {2020},
  url = {https://proceedings.neurips.cc/paper/2020/hash/c61f571dbd2fb949d3fe5ae1608dd48b-Abstract.html}
}

測試視圖與配對統計

對觀測視圖集 $\mathcal V$，訓練、驗證與測試切分分別為

$$ \mathcal V

\mathcal V_{\mathrm{train}} \cup \mathcal V_{\mathrm{val}} \cup \mathcal V_{\mathrm{test}}, \qquad \mathcal V_i\cap\mathcal V_j=\varnothing. $$

訓練圖像用於梯度更新，驗證圖像用於選擇權重、停止時刻或檢查點，測試圖像留作最終評價。對每個場景反覆根據測試 PSNR 調整閾值，測試視圖便實際參與了模型選擇；即使沒有對它反向傳播，人工選擇也會引入信息泄漏。

相機軌跡上的相鄰幀高度相關。固定步長抽取的測試幀通常仍靠近訓練軌跡，主要反映局部視角插值；按相機位置和觀察方向拉大間隔後，測試集會包含更明顯的視角變化。公開基準沿用社區常用切分，額外切分單獨成表，兩組結果各自對應一套視圖分佈。

令方向係數 $\eta_m=1$ 表示指標越大越好，$\eta_m=-1$ 表示越小越好。對同一個測試視圖 $v$，方法 B 相對 A 的指標差定義為

$$ \Delta_v

\eta_m\left(m_v^{(B)}-m_v^{(A)}\right). $$

平均差為

$$ \overline\Delta

\frac{1}{|\mathcal V_{\mathrm{test}}|} \sum_v\Delta_v. $$

按這個定義，$\Delta_v>0$ 表示 B 在視圖 $v$ 上更好。每個視圖自身的難度同時存在於 $m_v^{(A)}$ 和 $m_v^{(B)}$ 中，在差值裏相互抵消。兩個總體平均值 $\overline m^{(A)}$ 和 $\overline m^{(B)}$ 會隱去這層配對關係，也看不出收益分佈在多數視圖，還是集中於少量極端樣本。

不對 $\Delta_v$ 作高斯分佈假設時，可以對測試視圖做 bootstrap。每次有放回抽取 $|\mathcal V_{\mathrm{test}}|$ 個差值並計算均值，重複 $B$ 次得到 ${\overline\Delta^{(b)}}_{b=1}^B$，以其 $2.5%$ 與 $97.5%$ 分位數構造 $95%$ 置信區間：

$$ \overline\Delta \quad [q_{0.025},q_{0.975}]. $$

連續軌跡中的逐幀差值並不獨立，逐幀重抽樣會低估不確定性。按連續幀塊做 block bootstrap，或先計算場景均值再在場景層級重抽樣，可以保留一部分相關結構。場景數很少時，置信區間表示的仍是當前場景集內的波動。

弱紋理區域的統計

全圖誤差中，每個像素的權重相同，大面積平滑區域因像素數量較多而佔據主要比例。特定材質、弱紋理、邊界或高光區域可以用預先定義的掩碼 $M_v(\mathbf p)$ 單獨統計，其加權 MSE 為

$$ \operatorname{MSE}_M

\frac{ \sum_{v,\mathbf p} M_v(\mathbf p) \left| \widehat{\mathbf I}_v(\mathbf p)-\mathbf I_v(\mathbf p) \right|2^2 }{ 3\sum{v,\mathbf p}M_v(\mathbf p)+\varepsilon }. $$

對應的區域 PSNR 為

$$ \operatorname{PSNR}M =10\log{10} \frac{L^2}{\operatorname{MSE}_M}. $$

區域結果的可比性取決於掩碼是否獨立於待比較方法。若用某個方法的渲染深度或殘差定義「難區域」，不同方法實際上落在不同像素集上。由真值圖像、外部語義或邊緣檢測生成掩碼，或者預先固定幾何條件，可以讓所有結果共用同一統計區域。

圖像質量與幾何質量

對真值深度 $D^*$ 和預測深度 $\widehat D$，有效像素集為 $\Omega_D$，可計算

$$ \operatorname{RMSE}_D

\sqrt{ \frac{1}{|\Omega_D|} \sum_{\mathbf p\in\Omega_D} \left( \widehat D(\mathbf p)-D^*(\mathbf p) \right)^2 }. $$

全局尺度存在歧義時，尺度與偏置對齊會改變深度 RMSE 的含義。以最小二乘估計 $s$ 與 $b$ 可寫為

$$ (s^,b^)

\arg\min_{s,b} \sum_{\mathbf p\in\Omega_D} \left( s\widehat D(\mathbf p)+b-D^*(\mathbf p) \right)^2. $$

對點雲 $P$ 與真值表面 $Q$，對稱 Chamfer 距離為

$$ d_{\mathrm{CD}}(P,Q)

\frac{1}{|P|}\sum_{\mathbf x\in P} \min_{\mathbf y\in Q}|\mathbf x-\mathbf y|2 + \frac{1}{|Q|}\sum{\mathbf y\in Q} \min_{\mathbf x\in P}|\mathbf y-\mathbf x|_2. $$

第一項近似準確性，第二項近似完整性。只輸出少量高置信度點可以降低第一項，同時會損失表面覆蓋。準確性與完整性共同描述點集和真值表面的雙向接近程度，也可以進一步組成調和指標。

全場景平均 Chamfer 距離對這類局部失敗並不敏感。玻璃、高光與細薄結構通常只佔少量像素，局部孔洞會被大面積穩定表面稀釋；按材質類型和結構尺度拆分後，指標可以顯示這些區域的實際變化。圖像指標仍可能保持穩定，因為外觀分支可以在缺失幾何附近合成相近顏色。

消融中的主效應與交互項

以越大越好的指標 $y$ 為例，兩個模塊 A 與 B 的 $2\times2$ 消融包含四組結果：基線 $y_{00}$，只加 A 的 $y_{10}$，只加 B 的 $y_{01}$，以及同時加入兩者的 $y_{11}$。A 在 B 關閉時的效應為

$$ \Delta_A^{(B=0)}=y_{10}-y_{00}, $$

A 在 B 開啟時的效應為

$$ \Delta_A^{(B=1)}=y_{11}-y_{01}. $$

交互項為

$$ \Delta_{AB}

y_{11}-y_{10}-y_{01}+y_{00}. $$

$\Delta_{AB}=0$ 時，兩個模塊對該指標的影響在當前四組設置中近似可加；$\Delta_{AB}<0$ 對應收益重疊或相互干擾；$\Delta_{AB}>0$ 則顯示正向協同。它仍是特定數據、指標和訓練預算下的統計量，具體機制要結合中間輸出判斷。

《傳播式增密中的候選生成與幾何門控》引用的 GaussianPro 消融正好包含這四組結果。傳播與平面約束的 PSNR 單獨收益分別為 $0.95$ dB 和 $0.49$ dB，組合收益為 $1.15$ dB，交互項由此得到 $-0.29$ dB [@cheng2024gaussianpro]。負值表明兩項收益存在重疊；深度、法線、高斯數和候選接受統計可以進一步定位重疊發生在哪些區域與訓練階段。

四組結果共用數據切分、圖像預處理、相機位姿、初始點雲、訓練預算與評價實現，交互項才有一致的計算基礎。GaussianPro 額外列出不同 SfM 點雲下的 3DGS 基線，從傳播收益中分離了初始化變化。Color MLP 的對照固定高斯初始化與增密調度，本徵監督的對照固定偽標籤、置信掩碼和啟用時刻，差值便分別落在外觀參數化與監督信號上。

定量差值的空間分佈

均值 $\overline\Delta$ 把視圖間差異壓縮成一個數。將配對差值 ${\Delta_v}$ 排序後，中位數附近、改善分佈上分位和下分位的視圖分別對應典型結果、明顯收益與退化案例。預先固定這三個位置，也能減少先查看渲染圖再挑選樣本帶來的偏差。

視圖內部還可計算逐像素改善圖。以方法 B 相對 A 的 L1 殘差變化為例，

$$ G_v(\mathbf p)

\left| \widehat{\mathbf I}^{(A)}_v(\mathbf p)-\mathbf I_v(\mathbf p) \right|_1

\left| \widehat{\mathbf I}^{(B)}_v(\mathbf p)-\mathbf I_v(\mathbf p) \right|_1. $$

$G_v(\mathbf p)>0$ 的區域表示 B 的像素殘差較小。它與弱紋理掩碼、深度差和法線差並排後，可以顯示收益集中在大面積低頻區域、物體邊界，還是少量高對比像素。比較中的相機、裁切範圍、縮放倍數與色彩空間保持一致。

局部放大圖中的場景上下文用於辨認邊緣、遮擋關係和物體尺度。裁切過緊時，局部結構失去參照；裁切過鬆時，差異在網頁寬度下又會變得過小。保留主圖並附同比例放大框，可以兼顧兩種觀察尺度。

訓練過程與計算預算

記測試指標在訓練步 $t$ 為 $m(t)$。最終值 $m(T)$ 不含收斂速度、中途震蕩和後期退化的信息。對固定評價時刻 $t_1<\cdots<t_K$，歸一化曲線面積可以概括訓練期間的表現：

$$ \operatorname{AUC}_m

\frac{1}{t_K-t_1} \sum_{k=1}^{K-1} \frac{m(t_k)+m(t_{k+1})}{2} (t_{k+1}-t_k). $$

後期穩定性可在最後 $J$ 個檢查點上計算

$$ s_m

\sqrt{ \frac{1}{J-1} \sum_{k=K-J+1}^{K} \left(m(t_k)-\overline m_{\mathrm{tail}}\right)^2 }. $$

對 PSNR 與 SSIM，AUC 越大且 $s_m$ 越小，表示在當前評價頻率下收斂更快、後期波動更小；LPIPS 與誤差指標的方向相反。曲線平滑與隨機種子穩定性是兩個統計量，同一次訓練中的相鄰檢查點本身具有強相關性。

從整條測試曲線選擇最佳點，相當於使用了 $K$ 次測試反饋。由驗證集選定檢查點後，測試集只在該點評價一次；採用固定訓練預算時，各方法則共用同一預算定義。

終點質量與取得該結果的成本可以組成多目標向量：

$$ \mathbf y

\left[ -\operatorname{PSNR}, \operatorname{LPIPS}, d_{\mathrm{geom}}, t_{\mathrm{train}}, -\operatorname{FPS}, M_{\mathrm{model}} \right]. $$

對所有分量都以越小越好為約定。若方法 A 在每個分量上都不差於 B，且至少一項更好，則 A 支配 B。若 A 的圖像質量更高但訓練更慢、模型更大，兩者位於不同 Pareto 點，具體選擇取決於應用預算。

高斯數 $N$ 只是成本的一個代理量。真實渲染成本還受每個高斯的屏幕覆蓋範圍（footprint）、覆蓋 tile 數、深度排序實例和像素過度繪製影響。較少的高斯直接對應更緊湊的表示，FPS 仍以實測結果為準。

固定迭代數比較每一步優化帶來的收益，固定訓練時間比較端到端計算成本。傳播式增密增加候選生成和多視圖檢查，Color MLP 也改變每次前向與反向傳播的開銷，兩種預算下的排序可能不同。固定分辨率、硬件、預熱方式與數據寫回範圍後，FPS 即可與模型存儲和高斯數一起進入 Pareto 比較。

結語

新視角 PSNR、SSIM 和 LPIPS 衡量測試圖像與參考圖的接近程度；深度、法線與表面距離描述幾何誤差；反照率或光照分解則依賴相應真值。三類觀測可以同時改善，也會因外觀補償、尺度對齊和局部失敗而呈現不同趨勢。

對 3DGS 的增密、外觀解碼和幾何正則，配對差值給出收益在視圖間的分佈，區域指標和誤差圖指出變化發生的位置，消融交互項刻畫模塊組合，訓練曲線與 Pareto 向量補充取得結果所需的預算。這些觀測共同描述本次比較中的圖像質量、幾何變化、模塊作用與計算成本。

@inproceedings{cheng2024gaussianpro,
  author = {Cheng, Kai and Long, Xiaoxiao and Yang, Kaizhi and Yao, Yao and Yin, Wei and Ma, Yuexin and Wang, Wenping and Chen, Xuejin},
  title = {GaussianPro: 3D Gaussian Splatting with Progressive Propagation},
  booktitle = {Proceedings of the International Conference on Machine Learning},
  pages = {8123--8140},
  year = {2024}
}

@inproceedings{huang20242dgs,
  author = {Huang, Binbin and Yu, Zehao and Chen, Anpei and Geiger, Andreas and Gao, Shenghua},
  title = {2D Gaussian Splatting for Geometrically Accurate Radiance Fields},
  booktitle = {ACM SIGGRAPH 2024 Conference Papers},
  pages = {1--11},
  year = {2024}
}

@article{wang2004ssim,
  author = {Wang, Zhou and Bovik, Alan C. and Sheikh, Hamid R. and Simoncelli, Eero P.},
  title = {Image Quality Assessment: From Error Visibility to Structural Similarity},
  journal = {IEEE Transactions on Image Processing},
  volume = {13},
  number = {4},
  pages = {600--612},
  year = {2004}
}

@inproceedings{zhang2018lpips,
  author = {Zhang, Richard and Isola, Phillip and Efros, Alexei A. and Shechtman, Eli and Wang, Oliver},
  title = {The Unreasonable Effectiveness of Deep Features as a Perceptual Metric},
  booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages = {586--595},
  year = {2018}
}

這些模塊進入同一個訓練循環後，問題從「某個損失如何寫」變為「當多種誤差同時改變幾何、外觀和高斯集合時，殘差會被分配給誰」。加權求和只是目標函數的表面形式；實際優化還受殘差單位、有效像素數、梯度方向、參數共享範圍與離散增密調度影響。

參數、觀測與殘差空間

將場景參數分為幾何、外觀與共享解碼器三組：

$$ \boldsymbol\theta

\left[ \boldsymbol\theta_g^{\mathsf T}, \boldsymbol\theta_a^{\mathsf T}, \boldsymbol\theta_f^{\mathsf T} \right]^{\mathsf T}. $$

$\boldsymbol\theta_g$ 包含高斯中心、協方差與不透明度；$\boldsymbol\theta_a$ 包含逐高斯潛在特徵、顏色偏置或反照率代碼；$\boldsymbol\theta_f$ 是多個高斯共享的 Color MLP 參數。對視圖 $v$ 和像素 $\mathbf p$，多通道渲染器輸出

$$ \begin{aligned} \mathcal R_v(\boldsymbol\theta) ={}&\bigl( \widehat{\mathbf I}_v, \widehat{\mathbf A}_v, \widehat{\mathbf S}_v,\ &\widehat D_v, \widehat{\mathbf N}_v, W_v \bigr). \end{aligned} $$

$\widehat{\mathbf I}$ 是最終顏色，$\widehat{\mathbf A}$ 與 $\widehat{\mathbf S}$ 分別表示反照率與陰影/調制分支，$\widehat D$ 與 $\widehat{\mathbf N}$ 提供幾何觀測，$W$ 表示當前高斯對像素的前景解釋程度。同一個渲染結果可以生成多組殘差，但這些殘差不共享同一種統計含義。

圖像、本徵與幾何項

圖像保真項維持對真實觀測的擬合：

$$ \mathcal L_{\mathrm{rgb}}

(1-\lambda_s) \left| \widehat{\mathbf I}-\mathbf I \right|1 +\lambda_s\mathcal L{\mathrm{D\text{-}SSIM}}. $$

反照率偽標籤 $\mathbf A_v^*$ 在進入損失前已經做過跨視角校準，仍然只有局部可靠。記場景級質量為 $q_v$，像素級不確定性為 $u_v(\mathbf p)$，有效掩碼為 $m_v(\mathbf p)$，可定義

$$ \omega_v(\mathbf p) =q_v,m_v(\mathbf p) \exp!\left(-\frac{u_v(\mathbf p)}{\tau_u}\right), $$

$$ \mathcal L_A

\frac{ \sum_{\mathbf p} \omega_v(\mathbf p) \rho_A!\left( \widehat{\mathbf A}v(\mathbf p)-\mathbf A_v^*(\mathbf p) \right) }{ \sum{\mathbf p}\omega_v(\mathbf p)+\varepsilon }. $$

分母使損失不會因「這一幀只有少量可用像素」而系統性變小。$\rho_A$ 採用 Charbonnier 或 Huber 函數時，局部偽標籤失敗不會產生無界梯度。

幾何項可以包含深度分佈集中、渲染法線與深度法線一致，以及局部平面性 [@huang20242dgs; @turkulainen2025dnsplatter]。用 $w_i(\mathbf p)$ 表示射線上第 $i$ 個高斯的 Alpha 混合權重，歸一化權重為

$$ \widetilde w_i(\mathbf p)

\frac{w_i(\mathbf p)}{\sum_jw_j(\mathbf p)+\varepsilon}. $$

沿射線的深度方差為

$$ V_D(\mathbf p)

\sum_i\widetilde w_i \left(z_i-\overline z\right)^2, \qquad \overline z=\sum_i\widetilde w_i z_i. $$

深度分佈正則可寫成

$$ \mathcal R_{\mathrm{dist}}

\frac{ \sum_{\mathbf p}m_g(\mathbf p)V_D(\mathbf p) }{ \sum_{\mathbf p}m_g(\mathbf p)+\varepsilon }, $$

其中 $m_g$ 應排除半透明物體、深度不連續邊界與前景解釋權重過低的像素。法線一致項為

$$ \mathcal R_{\mathrm{normal}}

\frac{ \sum_{\mathbf p}m_g(\mathbf p) \left[ 1-\widehat{\mathbf N}(\mathbf p)^{\mathsf T} \mathbf N_D(\mathbf p) \right] }{ \sum_{\mathbf p}m_g(\mathbf p)+\varepsilon }. $$

$\mathbf N_D$ 由深度反投影後的局部切向叉積得到。當深度本身帶有噪聲，$\mathbf N_D$ 也會把高頻差分放大，因此它更適合與圖像邊緣權重和置信掩碼一起使用。

多任務目標與有效觀測

將各項放在一起，可以寫出一個通用的聯合目標：

$$ \begin{aligned} \mathcal L(\boldsymbol\theta,t) =;& \lambda_{\mathrm{rgb}}\mathcal L_{\mathrm{rgb}} +\eta_A(t)\lambda_A\mathcal L_A\ &+\eta_g(t) \left( \lambda_d\mathcal R_{\mathrm{dist}} +\lambda_n\mathcal R_{\mathrm{normal}} +\lambda_p\mathcal R_{\mathrm{plane}} \right)\ &+\eta_s(t) \left( \lambda_{\mathrm{id}}\mathcal R_{\mathrm{id}} +\lambda_{\mathrm{sm}}\mathcal R_{\mathrm{shading}} \right). \end{aligned} $$

$\eta_A(t)$、$\eta_g(t)$ 和 $\eta_s(t)$ 同時包含時間調度與數據可用性門控。當偽反照率不可用時，$\eta_A=0$；當像素處於深度邊界或前景覆蓋不足時，它不進入幾何項的有效集；當反照率分支缺少監督時，陰影恶性震蕩或任意分解的風險才需要 $\mathcal R_{\mathrm{id}}$ 與 $\mathcal R_{\mathrm{shading}}$。

用有效集改寫損失比「所有項永遠開啟，再慢慢調小權重」更清楚。權重表示可靠觀測之間的相對強度，門控則回答某個觀測是否屬於當前目標的定義域。

殘差尺度與歸一化

令第 $k$ 項損失是 $n_k$ 個有效殘差的平均：

$$ \mathcal L_k =\frac{1}{n_k} \sum_{j=1}^{n_k}\rho_k(r_{kj}). $$

即使各項都輸出一個標量，它們的數值範圍仍然受殘差單位與魯棒函數影響。RGB $L_1$ 的單位是顏色值，法線項是無量綱角度，深度方差又隨場景尺度的平方變化。若世界座標整體縮放 $s$，

$$ D\rightarrow sD, \qquad V_D\rightarrow s^2V_D. $$

固定 $\lambda_d$ 會讓同一個配置在不同場景尺度下實際強度不同。可以使用相對深度、以場景中位深度 $\widetilde D$ 歸一化，或將深度分佈項寫為

$$ \overline V_D

\frac{V_D}{\overline z^2+\varepsilon}. $$

歸一化還要覆蓋像素數。若一項按全圖求平均，另一項只對弱紋理掩碼求和，後者的有效面積會直接改變損失強度。因此，每項損失都應明確寫出「對誰求和」和「用什麼歸一化」。

梯度對齊與衝突

對共享參數子集 $\boldsymbol\theta_s$，第 $a$ 與第 $b$ 項損失的梯度為

$$ \mathbf g_a =\nabla_{\boldsymbol\theta_s}\mathcal L_a, \qquad \mathbf g_b =\nabla_{\boldsymbol\theta_s}\mathcal L_b. $$

它們的餘弦相似度為

$$ \rho_{ab}

\frac{ \mathbf g_a^{\mathsf T}\mathbf g_b }{ |\mathbf g_a|_2|\mathbf g_b|_2+\varepsilon }. $$

$\rho_{ab}>0$ 時，兩項在當前局部共同降低；$\rho_{ab}<0$ 時，某一項的下降方向會升高另一項。對反照率監督與 RGB 監督，衝突常出現在陰影、高光和曝光截斷區域；對圖像監督與平面幾何正則，衝突常出現在真實小結構與遮擋邊界。

若只查看損失數值，「反照率損失很小」並不代表它沒有主導某些參數。參數梯度還會經過解碼器雅可比、Alpha 混合權重與投影雅可比放大或縮小。實驗記錄中至少應分別統計幾何參數、逐高斯外觀參數和共享 MLP 的梯度范數。

外觀—幾何消元

將局部殘差線性化為

$$ \mathbf r \approx \mathbf r_0 +\mathbf J_g\delta\boldsymbol\theta_g +\mathbf J_a\delta\boldsymbol\theta_a. $$

對應的 Gauss–Newton 矩陣為

$$ \mathbf H= \begin{bmatrix} \mathbf H_{gg}&\mathbf H_{ga}\ \mathbf H_{ag}&\mathbf H_{aa} \end{bmatrix}

\begin{bmatrix} \mathbf J_g^{\mathsf T}\mathbf J_g& \mathbf J_g^{\mathsf T}\mathbf J_a\ \mathbf J_a^{\mathsf T}\mathbf J_g& \mathbf J_a^{\mathsf T}\mathbf J_a \end{bmatrix}. $$

消去外觀增量後，幾何的有效信息矩陣為

$$ \mathbf H_{g\mid a}

\mathbf H_{gg} -\mathbf H_{ga} \left(\mathbf H_{aa}+\lambda_a\mathbf I\right)^{-1} \mathbf H_{ag}. $$

當 Color MLP、逐高斯特徵與顏色偏置具有很強的殘差吸收能力，$\mathbf H_{aa}$ 的可用方向增加，消元項會減少留給幾何的曲率。Intrinsic 反照率監督和陰影正則可以限制外觀自由度，傳播式增密又為幾何增加新基底；這些模塊的作用不只體現在各自的損失值，還會改變其他殘差可以沿哪些參數方向下降。

不確定性加權

以同方差不確定性為例，多任務損失可寫成 [@kendall2018multitask]

$$ \mathcal L

\sum_k \left( \frac{1}{2\sigma_k^2}\mathcal L_k +\log\sigma_k \right). $$

學習到的 $\sigma_k$ 會降低噪聲較大任務的權重，$\log\sigma_k$ 防止所有權重縮到零。這個形式需要殘差對應明確的似然模型。本徵偽標籤中的結構性偏差、幾何邊界上的非高斯誤差，以及根本不可用的標籤，都不能只靠一個全局 $\sigma_k$ 解決。質量路由與像素掩碼仍然需要先定義有效集。

GradNorm 類方法會根據各任務梯度范數和相對訓練速度調整權重 [@chen2018gradnorm]。它能緩解「某一項因單位較大長期壓過其他項」，但梯度范數相等不代表方向相容。對 $\mathbf g_a^{\mathsf T}\mathbf g_b<0$ 的任務，PCGrad 類投影可寫成 [@yu2020pcgrad]

$$ \mathbf g_a^{\prime}

\mathbf g_a

\frac{ \mathbf g_a^{\mathsf T}\mathbf g_b }{ |\mathbf g_b|_2^2+\varepsilon } \mathbf g_b. $$

投影後 $\mathbf g_a^{\prime}$ 不再包含反對 $\mathbf g_b$ 的分量。在 3DGS 中，不建議對所有參數盲目做全局梯度手術：本徵項本來就應更直接影響外觀分支，法線項本來就應主要更新幾何。更有解釋性的做法是先限定各損失的參數作用域，再在真正共享的子空間中診斷衝突。

參數作用域

用 stop-gradient 或分組優化器可以把監督語義轉化為明確的梯度邊界。一個可用的參數—損失關係如下：

反照率項是否應更新幾何沒有固定答案。若偽標籤已做好對齊且邊界可靠，其梯度可以幫助幾何邊界；若偽標籤來自單幅二維模型並含有結構偏差，允許它直接拉動高斯中心會把二維幻覺固化成三維結構。可以用較小幾何學習率、延遲啟用，或只在多視圖一致區域放行這條梯度。

分階段調度

聯合目標在不同訓練階段的信任基礎不同。幾何未形成時，渲染反照率、深度和法線都只是初始化狀態的投影；訓練後期仍持續高頻增密，又會讓每次新增高斯重新擾動已穩定的外觀分解。一個通用調度可分為：

1. 表示熱身：以 RGB 保真建立可渲染的主要結構，反照率只做質量統計，幾何正則使用小權重。
2. 幾何形成：啟用深度分佈、法線一致與傳播式候選，並記錄候選接受率、剪枝率與高斯數量。
3. 外觀解耦：當基礎結構可靠後，依質量權重漸進啟用反照率監督，避免早期偽標籤將錯誤邊界寫入幾何。
4. 結構冷卻：停止或大幅降低增密頻率，讓已有高斯的幾何與外觀參數在固定結構上收斂。

漸進因子可以用平滑階躍表示。以反照率項為例，

$$ \eta_A(t)

q_v, \operatorname{smoothstep} \left( \frac{t-t_A}{\Delta t_A} \right). $$

$t_A$ 是啟用中心，$\Delta t_A$ 決定渡越寬度。直接從 $0$ 切換到固定大權重，會讓 Adam 的一階與二階動量同時面對殘差分佈突變，平滑啟用能夠降低這種瞬時震蕩。

訓練過程的梯度監測

單獨記錄總損失無法分辨「某項已收斂」與「某項已被其他參數屏蔽」。訓練日誌中更有用的診斷量包括：

聯合損失的價值不在於把更多項加到一行公式中。Intrinsic 把外觀殘差分配到材質與陰影語義，LE3D 式解碼器決定這些語義通過哪些參數表達，GaussianPro 式傳播在圖像梯度不足時提出新幾何，而深度與法線正則限制新舊圖元如何組成表面。只有將數據有效性、參數作用域、梯度尺度和結構調度同時寫清，這些模塊才能形成可診斷的優化管線。

《重建指標、消融證據與結論邊界》以測試視圖的配對統計、局部幾何、消融交互和計算預算整理這些變化的評價方式。

@article{kerbl2023gaussians,
  author = {Kerbl, Bernhard and Kopanas, Georgios and Leimkühler, Thomas and Drettakis, George},
  title = {3D Gaussian Splatting for Real-Time Radiance Field Rendering},
  journal = {ACM Transactions on Graphics},
  volume = {42},
  number = {4},
  pages = {139:1--139:14},
  year = {2023}
}

@inproceedings{jin2024lighting,
  author = {Jin, Xin and Jiao, Pengyi and Duan, Zheng-Peng and Yang, Xingchao and Li, Chongyi and Guo, Chun-Le and Ren, Bo},
  title = {Lighting Every Darkness with 3DGS: Fast Training and Real-Time Rendering for HDR View Synthesis},
  booktitle = {Advances in Neural Information Processing Systems},
  volume = {37},
  pages = {80191--80219},
  year = {2024}
}

@inproceedings{cheng2024gaussianpro,
  author = {Cheng, Kai and Long, Xiaoxiao and Yang, Kaizhi and Yao, Yao and Yin, Wei and Ma, Yuexin and Wang, Wenping and Chen, Xuejin},
  title = {GaussianPro: 3D Gaussian Splatting with Progressive Propagation},
  booktitle = {Proceedings of the International Conference on Machine Learning},
  pages = {8123--8140},
  year = {2024}
}

@inproceedings{huang20242dgs,
  author = {Huang, Binbin and Yu, Zehao and Chen, Anpei and Geiger, Andreas and Gao, Shenghua},
  title = {2D Gaussian Splatting for Geometrically Accurate Radiance Fields},
  booktitle = {ACM SIGGRAPH 2024 Conference Papers},
  pages = {1--11},
  year = {2024}
}

@inproceedings{turkulainen2025dnsplatter,
  author = {Turkulainen, Matias and Ren, Xin and Melekhov, Iaroslav and Seiskari, Otto and Rahtu, Esa and Kannala, Juho},
  title = {DN-Splatter: Depth and Normal Priors for Gaussian Splatting and Meshing},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages = {2421--2431},
  year = {2025}
}

@inproceedings{kendall2018multitask,
  author = {Kendall, Alex and Gal, Yarin and Cipolla, Roberto},
  title = {Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics},
  booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages = {7482--7491},
  year = {2018}
}

@inproceedings{chen2018gradnorm,
  author = {Chen, Zhao and Badrinarayanan, Vijay and Lee, Chen-Yu and Rabinovich, Andrew},
  title = {GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask Networks},
  booktitle = {Proceedings of the International Conference on Machine Learning},
  pages = {794--803},
  year = {2018}
}

@inproceedings{yu2020pcgrad,
  author = {Yu, Tianhe and Kumar, Saurabh and Gupta, Abhishek and Levine, Sergey and Hausman, Karol and Finn, Chelsea},
  title = {Gradient Surgery for Multi-Task Learning},
  booktitle = {Advances in Neural Information Processing Systems},
  volume = {33},
  pages = {5824--5836},
  year = {2020}
}

Office for Mac 的顯示語言會使用作業系統的主要語言設定。若要更改 Office for Mac 的顯示語言，需要為你的作業系統設定不同的主要語言：

在「Apple」選單中，選擇「系統偏好設定」>「語言與地區」

在「偏好語言」列表中，將所需語言拖曳到列表頂端，以將其設為主要語言。

如果未顯示所需語言，請點擊「+」符號新增。

可能需要重新啟動 Office 應用程式，才能讓此變更生效。

但出於習慣，我不能接受這種通過修改全局語言來實現需求的辦法。隨後通過檢索互聯網，發現了一個方便的方法，可以快速的切換Office套件的顯示語言。

以Word為例，如果需要將顯示語言切換至簡體中文，只需要在終端機中執行：

defaults write com.microsoft.Word AppleLanguages '("zh-cn")'

如果需要切換Powerpoint、Excel的顯示語言，僅需要將其中的Word更換成對應的軟體名稱就好。

需要切回默認顯示語言也同理，僅需要替換zh-cn為所需要的語言代碼即可。

《弱紋理場景中的幾何可觀測性》已經給出這個退化的局部形式：當圖像梯度趨近零，幾何位移對光度殘差的雅可比也會變弱。GaussianPro 將增密問題映射到有規則網格的二維圖像平面，先傳播深度與法線假設，再將通過多視圖檢查的像素反投影成新高斯 [@cheng2024gaussianpro]。這條路徑的核心是一個帶有拒絕機制的幾何提案器。

增密中的離散結構選擇

設當前場景含有 $N$ 個高斯，連續參數記為 $\Theta_N$。訓練問題可寫成

$$ \min_{N,\Theta_N} \sum_{v\in\mathcal V} \mathcal L_{\mathrm{img}} \left( \mathcal R(\Theta_N;\mathcal C_v), \mathbf I_v \right) +\beta,\Omega(\Theta_N,N). $$

Adam 在固定 $N$ 時更新 $\Theta_N$；增密、剪枝和不透明度重置會改變 $N$ 以及參數空間。$\Omega$ 可以表示模型規模、尺度畸變或幾何不連續的隱式代價。3DGS 以視空間梯度選擇 clone 或 split 的對象，用啟發式結構操作近似這個混合離散—連續優化。

對像素 $\mathbf p$ 的顏色殘差 $r(\mathbf p)$，高斯投影中心 $\mathbf u_i$ 接收到的梯度含有

$$ \frac{\partial r(\mathbf p)}{\partial \boldsymbol\mu_i}

\underbrace{ \frac{\partial r(\mathbf p)}{\partial \mathbf u_i} }{\text{圖像與覆蓋變化}} \underbrace{ \frac{\partial \mathbf u_i}{\partial \boldsymbol\mu_i} }{\text{透視投影}} . $$

第一項同時受圖像紋理、透射率、遮擋順序與當前高斯覆蓋影響。當某處沒有任何高斯投影，$\mathbf u_i$ 根本不存在；當區域近似常色，位移引起的光度變化又很小。因此，梯度門限擅長決定「已有圖元是否需要增加容量」，卻難以從空無一物的位置提出新幾何。

深度—法線假設的二維表示

對參考視圖 $r$ 中的像素 $\mathbf p=[u,v,1]^{\mathsf T}$，一個局部表面假設可用深度 $d$ 和參考相機下的單位法線 $\mathbf n$ 表示。像素對應的三維點為

$$ \mathbf X_r(d) =d,\mathbf K_r^{-1}\mathbf p. $$

當前高斯集合可以渲染出初始深度 $D_r^{(0)}$ 與法線 $\mathbf N_r^{(0)}$。若第 $i$ 個高斯的尺度為 $s_{i1},s_{i2},s_{i3}$，旋轉矩陣為 $\mathbf R_i$，可用最短主軸近似其法線：

$$ k_i=\arg\min_k s_{ik}, \qquad \mathbf n_i=\mathbf R_i\mathbf e_{k_i}. $$

經過符號統一後，深度與法線都能使用與 RGB 相同的 Alpha 混合權重生成二維圖。二維網格為每個像素提供固定數量的鄰域，比在不規則高斯集合中反復建立三維近鄰關係更適合做局部假設傳播。

平面單應性與光度代價

假設參考相機到源相機 $s$ 的相對位姿為 $(\mathbf R_{sr},\mathbf t_{sr})$，局部平面在參考相機座標中滿足

$$ \mathbf n^{\mathsf T}\mathbf X_r=d_\pi. $$

該平面在兩幅圖像之間誘導單應矩陣 [@hartley2004multiple]

$$ \mathbf H_{r\rightarrow s}(\mathbf n,d_\pi)

\mathbf K_s \left( \mathbf R_{sr} +\frac{\mathbf t_{sr}\mathbf n^{\mathsf T}}{d_\pi} \right) \mathbf K_r^{-1}. $$

右側分數的符號取決於平面方程與位姿定義；只要全文使用同一約定，單應映射的幾何含義不變。參考圖的局部圖像塊 $P_r(\mathbf p)$ 可以經 $\mathbf H_{r\rightarrow s}$ 畸變到源圖。常用的零均值歸一化互相關代價為

$$ C_{\mathrm{photo}}^{(s)}

\frac{1-\operatorname{ZNCC} \left(P_r,P_{s\leftarrow r}(\mathbf n,d_\pi)\right)}{2}. $$

ZNCC 消去圖像塊的加性偏置並歸一化對比度，對局部曝光變化比直接 RGB 差更穩定，仍然依賴圖像塊內有可區分的結構。完全常色的平面上，多個深度假設可以取得相近分數；傳播必須依靠鄰近的可觀測區域把幾何信息延伸進來。

鄰域傳播與假設選擇

記第 $t$ 輪傳播後像素 $\mathbf p$ 的假設為 $h_{\mathbf p}^{(t)}=(d_{\mathbf p}^{(t)},\mathbf n_{\mathbf p}^{(t)})$。從四鄰域或八鄰域收集候選 $\mathcal H_{\mathbf p}^{(t)}$，再以多視圖代價選擇

$$ h_{\mathbf p}^{(t+1)}

\arg\min_{h\in\mathcal H_{\mathbf p}^{(t)}} \sum_{s\in\mathcal V_{\mathbf p}} \omega_s(h), \rho!\left(C_{\mathrm{photo}}^{(s)}(h)\right). $$

$\rho$ 是截斷或 Huber 類的魯棒函數，$\omega_s$ 由視角、遮擋、投影範圍與源視圖可用性決定。這一步與 PatchMatch 類多視圖立體匹配共享「鄰域傳播 + 隨機擾動 + 代價競爭」的結構 [@bleyer2011patchmatch]。從優化角度看，它以有限候選近似求解每個像素的非凸深度—法線問題，並用空間鄰近性限制搜索範圍。

傳播次數不宜無限增加。若一個錯誤平面假設在局部代價上占優，過多輪傳播會擴大其空間範圍。因此，傳播只負責提案，寫入三維場景前還需要更嚴格的多視圖門控。

多視圖幾何一致性

參考像素的候選三維點為 $\mathbf X_r$。將它投影到源視圖 $s$，得到 $\mathbf p_s=\pi_s(\mathbf X_r)$ 與預測深度 $d_{r\rightarrow s}$。若源視圖的當前深度為 $D_s(\mathbf p_s)$，可定義相對深度誤差

$$ e_d^{(s)}

\frac{ \left|d_{r\rightarrow s}-D_s(\mathbf p_s)\right| }{d_{r\rightarrow s}+D_s(\mathbf p_s)+\varepsilon}. $$

將 $\mathbf p_s$ 依源視圖深度反投影，再投回參考視圖，可得到循環重投影誤差

$$ e_p^{(s)}

\left| \mathbf p- \pi_r!\left( \mathbf T_{rs} \pi_s^{-1}(\mathbf p_s,D_s(\mathbf p_s)) \right) \right|_2. $$

法線在坐標轉換後的角度誤差為

$$ e_n^{(s)}

1- \left| (\mathbf R_{sr}\mathbf n_r)^{\mathsf T} \mathbf n_s(\mathbf p_s) \right|. $$

絕對值消去法線朝向的二義性，但在有明確前後關係的渲染器中，更合理的做法是先統一法線朝向再比較。源視圖通過檢查的條件可寫成

$$ v_s

\mathbb I \left[ e_d^{(s)}<\tau_d, \ e_p^{(s)}<\tau_p, \ e_n^{(s)}<\tau_n \right]. $$

只用單一最佳源視圖會放大偶然匹配，多視圖選擇需要同時考慮視角、遮擋與幾何支持 [@schonberger2016pixelwise]。對有效集合 $\mathcal V_k={s:v_s=1}$，可組織魯棒幾何代價

$$ C_{\mathrm{geom}}^{(k)}

\operatorname{median}_{s\in\mathcal V_k} \left( e_d^{(s)}+\lambda_p e_p^{(s)}+\lambda_n e_n^{(s)} \right). $$

中位數允許少量視圖因遮擋或反射而失敗，同時要求 $|\mathcal V_k|\ge n_{\min}$ 才能防止「只有一個支持者」的脆弱候選進入場景。

候選接受門控

傳播深度 $d_k^{\mathrm{prop}}$ 與當前渲染深度 $d_k^{\mathrm{rend}}$ 之間還需要有足夠大的差異：

$$ \Delta d_k

\frac{ \left|d_k^{\mathrm{prop}}-d_k^{\mathrm{rend}}\right| }{d_k^{\mathrm{rend}}+\varepsilon}. $$

$\Delta d_k$ 太小時，當前高斯已經解釋了該區域，新增圖元只會提高冗餘。綜合弱紋理掩碼、多視圖支持、光度代價與幾何代價，最終門控可寫成

$$ \delta_k= \mathbb I \left[ m_{\mathrm{weak}}(\mathbf p_k)=1, \ |\mathcal V_k|\ge n_{\min}, \ C_{\mathrm{photo}}^{(k)}<\tau_c, \ C_{\mathrm{geom}}^{(k)}<\tau_g, \ \Delta d_k>\tau_{\Delta d} \right]. $$

硬門控便於調試，閾值附近的候選會出現不連續翻轉。可以進一步構造連續置信度

$$ q_k

\sigma!\left(a_v(|\mathcal V_k|-n_{\min})\right) \sigma!\left(a_d(\Delta d_k-\tau_{\Delta d})\right) \exp!\left( -\frac{C_{\mathrm{photo}}^{(k)}}{\tau_c} -\frac{C_{\mathrm{geom}}^{(k)}}{\tau_g} \right), $$

再在每輪中選擇 $q_k$ 最高的固定數量候選。這個改寫把「門限」轉化為「置信度排序 + 增長預算」，可以避免場景尺度或圖像數量改變時高斯數量突然失控。

候選高斯的三維初始化

通過門控的像素先反投影為

$$ \boldsymbol\mu_k

\mathbf T_{wr} \begin{bmatrix} d_k\mathbf K_r^{-1}\mathbf p_k\ 1 \end{bmatrix}. $$

令世界座標下的法線為 $\mathbf n_k$，在切平面內構造正交基 $\mathbf t_{k1},\mathbf t_{k2}$，則旋轉矩陣可取

$$ \mathbf R_k= \begin{bmatrix} \mathbf t_{k1}&\mathbf t_{k2}&\mathbf n_k \end{bmatrix}. $$

單像素在深度 $d_k$ 處對應的世界尺度約為 $d_k/f$。若局部支持區半徑為 $r_p$ 個像素，初始切向尺度可設為

$$ s_{k1}=s_{k2}=\eta_t\frac{r_p d_k}{f}, \qquad s_{k3}=\eta_n\min(s_{k1},s_{k2}), \quad \eta_n\ll1. $$

扁平的初始協方差將新高斯限制在候選平面附近，同時保留後續優化尺度與旋轉的空間。顏色、潛在特徵或偏置可由鄰近可見高斯的透射率權重均值初始化；若直接複製單個最近高斯，遮擋邊界兩側的外觀更容易被混合。

在插入前還要檢查現有高斯。以局部協方差 $\boldsymbol\Sigma_j$ 定義馬氏距離

$$ d_M^2(k,j)

(\boldsymbol\mu_k-\boldsymbol\mu_j)^{\mathsf T} \boldsymbol\Sigma_j^{-1} (\boldsymbol\mu_k-\boldsymbol\mu_j). $$

若候選與某個現有高斯的 $d_M^2$ 很小，且法線與外觀都一致，更合適的操作是更新現有高斯的置信統計，而非再創建一個幾乎重疊的圖元。

調度、預算與穩定性

傳播式增密不適合每次梯度更新都執行。深度與法線在訓練早期還不穩定，頻繁插入候選會把瞬時誤差固化為新高斯。一個可解釋的調度策略包含四個時間尺度：

設每輪候選數為 $M_t$，接受上限為 $B_t$，則增長率為

$$ g_t=\frac{\min(M_t,B_t)}{N_t}. $$

保持 $g_t$ 隨訓練逐步衰減，能夠防止後期因少量高頻殘差持續增加模型規模。增長預算還應與剪枝率、可見高斯數、tile 實例數與顯存佔用一起記錄；高斯總數相同時，更大的屏幕 footprint 仍然可能帶來更高的光柵化成本。

傳播與平面約束的消融結果

GaussianPro 的消融實驗將傳播和平面約束分開。在 Waymo 實驗中，基線的 PSNR 為 $33.53$ dB；只使用傳播時為 $34.48$ dB，再加入平面約束後為 $34.68$ dB。對應的 SSIM 從 $0.938$ 增至 $0.949$，LPIPS 從 $0.226$ 降至 $0.191$ [@cheng2024gaussianpro]。

從消融表還能計算交互項。記 PSNR 基線為 $y_0$，傳播、平面約束和兩者同時使用的結果分別為 $y_P,y_L,y_{PL}$，則

$$ \Delta_{\mathrm{int}} =y_{PL}-y_P-y_L+y_0 =34.68-34.48-34.02+33.53 =-0.29\ \mathrm{dB}. $$

負交互項表明兩個模塊修正了部分重疊誤差，不適合把單獨增益直接相加當成組合收益。它也提示一個重要實驗要求：傳播提供了新結構，平面約束改變了已有結構的參數，兩者要同時用候選接受率、深度/法線指標與 RGB 指標判斷。

傳播式增密的失效情形

傳播依賴已有局部表面作為種子，並默認鄰域中存在可延伸的幾何結構。以下情況需要額外診斷：

• 完全無紋理且沒有邊界錨點的大平面，局部圖像塊不能區分不同深度假設。
• 鏡面、透明物體和強高光破壞跨視圖光度一致性，ZNCC 也無法將反射運動轉化為幾何證據。
• 細桿、植被和穿孔結構不滿足局部單平面假設，鄰域傳播會把前後層合併。
• 相機位姿誤差會同時升高光度與幾何代價，若放寬閾值強行保留候選，錯位會被寫入高斯集合。
• 遮擋邊界的深度與法線本來就不連續，平面傳播與平滑約束應在邊界兩側分別運行。
• 外觀模型容量過強時，錯誤幾何仍可被顏色偏置或解碼器補償，候選生成與外觀監督之間需要共同調度。

傳播式增密為弱紋理區域增加了 RGB 中心梯度之外的結構信號，但它沒有消除幾何歧義。完整路徑包含假設生成、多視圖檢查、反投影初始化、去重、增長預算與後期剪枝；其中任意一步過於寬鬆，都會把二維匹配中的偶然最優轉化為長期存在的三維圖元。

《本徵、外觀與幾何監督的聯合優化》將繼續處理這條關係：傳播改變高斯集合，Intrinsic 偽監督約束材質，LE3D 式顏色解碼器分配外觀殘差，幾種信號並不天然具有相容的尺度和梯度方向。

@article{kerbl2023gaussians,
  author = {Kerbl, Bernhard and Kopanas, Georgios and Leimkühler, Thomas and Drettakis, George},
  title = {3D Gaussian Splatting for Real-Time Radiance Field Rendering},
  journal = {ACM Transactions on Graphics},
  volume = {42},
  number = {4},
  pages = {139:1--139:14},
  year = {2023}
}

@inproceedings{cheng2024gaussianpro,
  author = {Cheng, Kai and Long, Xiaoxiao and Yang, Kaizhi and Yao, Yao and Yin, Wei and Ma, Yuexin and Wang, Wenping and Chen, Xuejin},
  title = {GaussianPro: 3D Gaussian Splatting with Progressive Propagation},
  booktitle = {Proceedings of the International Conference on Machine Learning},
  pages = {8123--8140},
  year = {2024}
}

@inproceedings{bleyer2011patchmatch,
  author = {Bleyer, Michael and Rhemann, Christoph and Rother, Carsten},
  title = {PatchMatch Stereo---Stereo Matching with Slanted Support Windows},
  booktitle = {Proceedings of the British Machine Vision Conference},
  pages = {14.1--14.11},
  year = {2011}
}

@inproceedings{schonberger2016pixelwise,
  author = {Schönberger, Johannes L. and Zheng, Enliang and Frahm, Jan-Michael and Pollefeys, Marc},
  title = {Pixelwise View Selection for Unstructured Multi-View Stereo},
  booktitle = {Proceedings of the European Conference on Computer Vision},
  pages = {501--518},
  year = {2016}
}

@book{hartley2004multiple,
  author = {Hartley, Richard and Zisserman, Andrew},
  title = {Multiple View Geometry in Computer Vision},
  edition = {2},
  publisher = {Cambridge University Press},
  year = {2004}
}

球諧顏色的解析結構

3DGS 為第 $i$ 個高斯保存球諧（Spherical Harmonics，SH）係數，用觀察方向 $\mathbf d\in\mathbb S^2$ 解碼顏色 [@kerbl2023gaussians]。對顏色通道 $c$，階數為 $L$ 的展開為

$$ c_{i,c}(\mathbf d) =\sum_{l=0}^{L} \sum_{m=-l}^{l} k_{i,c,lm}Y_l^m(\mathbf d). $$

令

$$ \mathbf y_L(\mathbf d) =\begin{bmatrix} Y_0^0(\mathbf d)&Y_1^{-1}(\mathbf d)&\cdots&Y_L^L(\mathbf d) \end{bmatrix}^{\mathsf T}, $$

則三通道顏色可以緊湊寫成

$$ \mathbf c_i(\mathbf d) =\mathbf K_i\mathbf y_L(\mathbf d), \qquad \mathbf K_i\in\mathbb R^{3\times(L+1)^2}. $$

固定方向後，顏色對係數是線性的：

$$ \frac{\partial\mathbf c_i}{\partial\operatorname{vec}(\mathbf K_i)} =\mathbf y_L^{\mathsf T}(\mathbf d)\otimes\mathbf I_3. $$

這一結構帶來穩定而便宜的前向計算，也使每個高斯的外觀參數相互獨立。$L=0$ 只有常量顏色；增加階數後，基函數數量按 $(L+1)^2$ 增長。3DGS 常逐步啟用更高階球諧，使優化先建立近似漫反射基色，再學習視角相關分量。

球諧函數是定義在完整球面上的正交基，但訓練相機只覆蓋有限方向集合 ${\mathbf d_v}$。對單個高斯，把這些方向上的基向量堆疊成

$$ \mathbf Y_i= \begin{bmatrix} \mathbf y_L^{\mathsf T}(\mathbf d_1)\ \vdots\ \mathbf y_L^{\mathsf T}(\mathbf d_M) \end{bmatrix}. $$

可辨識的係數方向由 $\operatorname{rank}(\mathbf Y_i)$ 決定。若高斯僅從很窄的視角範圍可見，某些高階基在採樣方向上近似相關，$\mathbf Y_i^{\mathsf T}\mathbf Y_i$ 會病態。訓練視圖顏色可以擬合得很好，未觀測方向卻可能出現振盪或顏色突變。解析基的正交性只在連續完整球面積分下成立，不保證有限相機採樣下的數值正交。

共享神經解碼器的參數組織

無球諧表示通常讓每個高斯保存低維潛在特徵 $\mathbf f_i\in\mathbb R^D$ 與可選顏色偏置 $\mathbf b_i$，再由共享網絡 $F_{\boldsymbol\phi}$ 根據特徵和視角生成顏色：

$$ \mathbf c_i(\mathbf d) =g!\left( F_{\boldsymbol\phi}(\mathbf f_i,\gamma(\mathbf d)) +\mathbf b_i \right). $$

$\gamma$ 是方向編碼，$g$ 決定輸出域。LDR 顏色可以使用 sigmoid 或有界映射；在線性 HDR 域中，上界會截斷亮度，LE3D 使用指數輸出

$$ \mathbf c_i =\exp!\left( F_{\boldsymbol\phi}(\mathbf f_i,\mathbf v) +\mathbf b_i \right), $$

並用多視圖投影顏色的對數均值初始化偏置，使初始輸出接近觀測亮度 [@jin2024lighting]。

共享網絡把外觀先驗從“每個高斯獨立的一組解析係數”改為“每個高斯代碼與全局解碼函數”。其參數量近似為

$$ P_{\mathrm{MLP}} =ND+N_b+P_{\boldsymbol\phi}, $$

球諧參數量則為

$$ P_{\mathrm{SH}} =3N(L+1)^2. $$

當 $N$ 很大且 $D<3(L+1)^2$ 時，潛在代碼可以降低逐高斯存儲；但共享網絡的計算成本、緩存訪問和批處理方式會進入渲染開銷。參數更少也不等於表達更弱：$F_{\boldsymbol\phi}$ 在全部高斯之間共享統計規律，能夠用非線性組合複用容量。

解碼與 Alpha 混合的次序

“先解碼再混合”和“先混合再解碼”經常都被稱為顏色 MLP，二者在非線性網絡下並不等價。

逐高斯解碼先得到

$$ \mathbf c_i=F(\mathbf f_i,\mathbf d_i), $$

再執行 Alpha 混合：

$$ \mathbf C_{\mathrm{pre}}(p) =\sum_iw_i(p)\mathbf c_i. $$

像素級解碼先聚合特徵

$$ \overline{\mathbf f}(p) =\frac{\sum_iw_i(p)\mathbf f_i}{W(p)+\epsilon}, \qquad W(p)=\sum_iw_i(p), $$

再生成顏色：

$$ \mathbf C_{\mathrm{post}}(p) =F(\overline{\mathbf f}(p),\mathbf d_p). $$

若 $F$ 對特徵是仿射函數，且權重已歸一化，則兩者可以相等。對一般非線性網絡，令

$$ \widetilde w_i=\frac{w_i}{W}, \qquad \overline{\mathbf f}=\sum_i\widetilde w_i\mathbf f_i, $$

在 $\overline{\mathbf f}$ 附近作二階展開：

$$ F(\mathbf f_i) \approx F(\overline{\mathbf f}) +\mathbf J_F(\mathbf f_i-\overline{\mathbf f}) +\frac12 (\mathbf f_i-\overline{\mathbf f})^{\mathsf T} \mathbf H_F (\mathbf f_i-\overline{\mathbf f}). $$

按 $\widetilde w_i$ 求和後，一階項消失：

$$ \sum_i\widetilde w_iF(\mathbf f_i) -F(\overline{\mathbf f}) \approx \frac12\sum_i\widetilde w_i (\mathbf f_i-\overline{\mathbf f})^{\mathsf T} \mathbf H_F (\mathbf f_i-\overline{\mathbf f}). $$

差異由特徵在射線上的方差和網絡曲率共同決定。單一表面、權重高度集中時，兩者接近；遮擋邊界、半透明區域或多層高斯混合時，像素級解碼會先把不同表面的特徵混在一起，可能生成任何一個高斯都沒有的顏色。逐高斯解碼保持圖元語義，代價是對每個可見高斯執行網絡或查表。

視角方向的導數

球諧顏色的方向導數為

$$ \frac{\partial\mathbf c_i}{\partial\mathbf d} =\mathbf K_i \frac{\partial\mathbf y_L(\mathbf d)}{\partial\mathbf d}. $$

最高階數限制了方向頻率，也使導數具有明確的解析結構。神經解碼器則有

$$ \frac{\partial\mathbf c_i}{\partial\mathbf d} =\mathbf J_{\mathrm{out}} \frac{\partial F_{\boldsymbol\phi}} {\partial\gamma} \frac{\partial\gamma(\mathbf d)}{\partial\mathbf d}. $$

若方向編碼包含高頻正餘弦，$\partial\gamma/\partial\mathbf d$ 會隨頻率增大；網絡能夠表示更尖銳的視角變化，也更容易在訓練方向之間出現高頻振盪。對鏡面高光這可能是必要容量，對漫反射牆面則可能成為記憶相機方向的捷徑。

方向本身由相機中心 $\mathbf o_v$ 與高斯位置 $\boldsymbol\mu_i$ 決定：

$$ \mathbf d_{iv} =\frac{\boldsymbol\mu_i-\mathbf o_v} {|\boldsymbol\mu_i-\mathbf o_v|_2}. $$

令 $r=|\boldsymbol\mu_i-\mathbf o_v|_2$，則

$$ \frac{\partial\mathbf d_{iv}} {\partial\boldsymbol\mu_i} =\frac{1}{r} \left( \mathbf I_3-\mathbf d_{iv}\mathbf d_{iv}^{\mathsf T} \right). $$

因此，顏色對位置還存在一條經過觀察方向的路徑：

$$ \frac{\partial\mathbf c_i} {\partial\boldsymbol\mu_i} =\frac{\partial\mathbf c_i}{\partial\mathbf d_{iv}} \frac{\partial\mathbf d_{iv}}{\partial\boldsymbol\mu_i}. $$

當神經外觀對方向非常敏感時，移動高斯不僅改變投影權重，也會顯著改變解碼顏色。幾何梯度由此混入外觀方向導數，位置更新更難單獨解釋為投影對齊。

外觀容量與幾何補償

設所有像素殘差堆疊為

$$ \mathbf r =\mathbf I_{\mathrm{render}}(\boldsymbol\theta_g,\boldsymbol\theta_a) -\mathbf I_{\mathrm{gt}}, $$

$\boldsymbol\theta_g$ 是幾何參數，$\boldsymbol\theta_a$ 是外觀參數。局部線性化得到

$$ \mathbf r(\Delta\boldsymbol\theta_g,\Delta\boldsymbol\theta_a) \approx \mathbf r_0 +\mathbf J_g\Delta\boldsymbol\theta_g +\mathbf J_a\Delta\boldsymbol\theta_a. $$

消去外觀增量後的幾何正規矩陣為

$$ \mathbf H_{g\mid a} =\mathbf J_g^{\mathsf T} \left[ \mathbf I -\mathbf J_a (\mathbf J_a^{\mathsf T}\mathbf J_a+\lambda_a\mathbf I)^{-1} \mathbf J_a^{\mathsf T} \right] \mathbf J_g. $$

中括號把可以由外觀子空間解釋的殘差方向投影出去。$\operatorname{col}(\mathbf J_a)$ 越大，落到幾何上的有效信息越少。增加 MLP 寬度、潛在特徵維度或方向編碼頻率，可能提高訓練圖像擬合，同時降低某些幾何方向的條件數。外觀容量與幾何質量之間不存在單調關係。

一個更直觀的標量例子是帶顏色斜率的邊緣。設投影位置誤差為 $\delta u$，圖像梯度為 $I_u$，外觀偏置為 $\delta b$，殘差近似為

$$ r\approx I_u\delta u+\delta b. $$

若 $\delta b$ 可自由調整，則取 $\delta b=-I_u\delta u$ 可以完全抵消位置誤差。給偏置加入 $\lambda_b\delta b^2$ 後，消去 $\delta b$ 得到位置方向的有效曲率

$$ H_{u\mid b} =I_u^2\frac{\lambda_b}{1+lambda_b}. $$

$\lambda_b\rightarrow0$ 時，幾何曲率趨近於零；強約束外觀後，位置才重新受到圖像證據約束。這個推導解釋了為何獨立顏色偏置既能加速初期擬合，也可能在弱紋理區域長期掩蓋幾何偏移。

反照率與陰影分支的規範

若神經解碼器輸出反照率與陰影

$$ \mathbf A_i =\sigma(F_A(\mathbf f_i,\mathbf d)), \qquad \mathbf S_i =g_S(F_S(\mathbf f_i,\mathbf d)), $$

並組合為

$$ \mathbf c_i=\mathbf A_i\odot\mathbf S_i+mathbf b_i, $$

則本徵分解的尺度自由度直接進入三維外觀表示。對正數 $\mathbf k$，

$$ (\mathbf A_i,\mathbf S_i) \mapsto (\mathbf k\odot\mathbf A_i, \mathbf S_i\oslash\mathbf k) $$

保持顏色不變；若還有自由偏置 $\mathbf b_i$，加性誤差也能被第三條路徑吸收。僅用 RGB 重建損失無法讓分支獲得穩定語義。

跨視角反照率偽監督為 $\mathbf A$ 選擇規範，陰影正則限制 $\mathbf S$ 的空間與時間變化，偏置容量則應保持受控。聯合目標可以寫成

$$ \mathcal L =\mathcal L_{\mathrm{rgb}} +w_A\mathcal L_A +\lambda_S\mathcal R_S +\lambda_b\mathcal R_b. $$

如果 $w_A=0$，反照率與陰影的變量名不保證任何分解；如果 $w_A$ 過大且偽標籤有偏，錯誤材質會反向污染共享特徵。第 6 篇中的質量路由恰好控制這一入口。

對像素級特徵解碼，還需決定反照率偽標籤監督的是

$$ F_A!\left( \sum_i\widetilde w_i\mathbf f_i \right) $$

還是

$$ \sum_i\widetilde w_iF_A(\mathbf f_i). $$

兩者的差異仍由網絡曲率和射線特徵方差決定。在遮擋邊界處，前者容易把前後表面的材質代碼混合，後者更接近可見圖元的反照率合成；但後者計算更重，並要求每個圖元分支具有明確含義。

初始化、複製與分裂

神經顏色表示引入了新的結構演化問題。高斯複製或分裂時，子高斯通常繼承父高斯的 $\mathbf f_i$ 與 $\mathbf b_i$。這樣能保持渲染連續，但兩個子高斯初始外觀完全相同。若後續可見視角也相似，它們可能長期無法分化。

設兩個子高斯特徵為

$$ \mathbf f_{i_1}=\mathbf f_i+\boldsymbol\epsilon, \qquad \mathbf f_{i_2}=\mathbf f_i-\boldsymbol\epsilon. $$

對稱小擾動可打破完全相同的狀態，但 $\boldsymbol\epsilon$ 過大會在緻密化瞬間造成顏色閃爍。另一種方式是複製特徵而改變幾何，讓可見性差異自然產生不同梯度。選擇取決於解碼器是否容易出現特徵塌縮，以及渲染連續性要求。

偏置初始化同樣影響優化。若使用觀測顏色均值初始化 $\mathbf b_i$，網絡殘差從接近零開始，能夠快速得到合理圖像；但觀測均值包含陰影和曝光，偏置可能先佔據本應由反照率或陰影分支解釋的成分。初始化提供的是優化起點，不應被當作材質估計。

解析基與神經解碼的取捨

球諧適合受限頻率的視角相關外觀，計算路徑簡單；神經解碼適合線性 HDR、複雜相機響應或需要顯式反照率/陰影接口的場景。選擇外觀模型時，至少要同時報告訓練視圖擬合、新視角穩定性、方向導數、幾何指標與運行開銷。只比較 PSNR 會把外觀容量和幾何質量混為一個結論。

無球諧表示的真正變化不在於刪除一組係數，而在於重新分配場景級與圖元級參數。共享網絡可以建立跨高斯外觀先驗，潛在代碼可以提供更靈活的輸入，反照率分支也能接收經過路由的偽監督；相應代價是非線性混合、規範自由度和更大的外觀捷徑空間。《傳播式增密中的候選生成與幾何門控》將局部 RGB 梯度不足時的結構增長問題獨立展開，並保留這條外觀消元視角。

@article{kerbl2023gaussians,
  author = {Kerbl, Bernhard and Kopanas, Georgios and Leimkühler, Thomas and Drettakis, George},
  title = {3D Gaussian Splatting for Real-Time Radiance Field Rendering},
  journal = {ACM Transactions on Graphics},
  volume = {42},
  number = {4},
  pages = {139:1--139:14},
  year = {2023}
}

@inproceedings{jin2024lighting,
  author = {Jin, Xin and Jiao, Pengyi and Duan, Zheng-Peng and Yang, Xingchao and Li, Chongyi and Guo, Chun-Le and Ren, Bo},
  title = {Lighting Every Darkness with 3DGS: Fast Training and Real-Time Rendering for HDR View Synthesis},
  booktitle = {Advances in Neural Information Processing Systems},
  volume = {37},
  pages = {80191--80219},
  year = {2024}
}

跨視角處理需要完成兩層工作。第一層是校準：估計可以由低維變換解釋的全局或逐通道漂移。第二層是決策：判斷校準後的殘差來自正常噪聲、局部失敗還是錯誤對應，並據此決定監督是否進入訓練。兩層不能合併為一個誤差閾值，因為一個自由度過高的校準模型也能降低殘差，卻不一定提高偽標籤質量。

多視圖重疊區域

設第 $i$ 個視角的反照率預測為 $\widehat{\mathbf A}_i$，參考視角為 $r$。像素 $p$ 通過當前深度 $D_i(p)$ 反投影到三維：

$$ \mathbf X_i(p) =\mathbf R_i^{\mathsf T} \left(D_i(p)\mathbf K_i^{-1}\widetilde p-\mathbf t_i\right), $$

再投影到參考視角：

$$ q_{ir}(p) =\pi!\left( \mathbf K_r (\mathbf R_r\mathbf X_i(p)+\mathbf t_r) \right). $$

這裏的 $\widetilde p$ 是齊次像素座標。只有同時滿足下列條件的像素才能進入重疊域 $\Omega_{ir}$：投影位於參考圖像內部，兩側深度在容差內一致，視線未被遮擋，反照率預測有效，且像素不處於明顯飽和或動態區域。可以寫成

$$ m_{ir}(p) =m_{\mathrm{in}}(p)\cdot m_{\mathrm{depth}}(p)\cdot m_{\mathrm{visible}}(p)\cdot m_{\mathrm{valid}}(p) \in{0,1}, $$

$$ \Omega_{ir}={p:m_{ir}(p)=1}, $$

其中乘號表示多個二值條件的合取。

深度一致性常用相對形式：

$$ m_{\mathrm{depth}}(p) =\mathbb{I}!\left[ \frac{|z_r(\mathbf X_i(p))-D_r(q_{ir}(p))|} {D_r(q_{ir}(p))+\epsilon}<\tau_z \right]. $$

若幾何本身尚不穩定，$\Omega_{ir}$ 會繼承深度誤差。反照率校準因此不應在訓練最早期由任意像素對應驅動；更穩妥的做法是使用可靠的初始幾何、稀疏匹配或經過前後向檢查的對應，並保留重疊率作為質量變量。

逐通道仿射校準

本徵分解中的尺度規範、相機曝光與白平衡通常表現為低維色彩漂移。對顏色通道 $c$，用仿射模型

$$ A^{*}{i,c}(p) =a{i,c}\widehat A_{i,c}(p)+b_{i,c} $$

將第 $i$ 幀映射到參考域。忽略魯棒損失時，加權最小二乘問題為

$$ \min_{a,b} \sum_{p\in\Omega_{ir}} \omega_p \left(a x_p+b-y_p\right)^2, $$

其中

$$ x_p=\widehat A_{i,c}(p), \qquad y_p=\widehat A_{r,c}(q_{ir}(p)). $$

令

$$ \overline x_w=\frac{\sum_p\omega_px_p}{\sum_p\omega_p}, \qquad \overline y_w=\frac{\sum_p\omega_py_p}{\sum_p\omega_p}, $$

閉式解為

$$ a^{} =\frac{ \sum_p\omega_p(x_p-\overline x_w)(y_p-\overline y_w) }{ \sum_p\omega_p(x_p-\overline x_w)^2 }, \qquad b^{}=\overline y_w-a^{*}\overline x_w. $$

分母是源預測在有效重疊域內的加權方差。當場景幾乎只有均勻白牆時，

$$ \sum_p\omega_p(x_p-\overline x_w)^2\approx0, $$

尺度 $a$ 變得不可辨識；這時很小的像素噪聲也會導致很大的尺度變化。給分母加 $\epsilon$ 只能避免數值除零，不能恢復缺失的信息。工程上需要限制 $a$ 與 $b$ 的合理範圍、增加先驗，或者在色彩跨度不足時退化為僅估計偏置/尺度。

將問題寫成矩陣形式

$$ \mathbf y=\mathbf X\boldsymbol\beta+\boldsymbol\epsilon, \qquad \mathbf X= \begin{bmatrix} x_1&1\ \vdots&\vdots\ x_n&1 \end{bmatrix}, \qquad \boldsymbol\beta= \begin{bmatrix}a\b\end{bmatrix}, $$

若殘差方差近似為 $\sigma^2$，參數協方差為

$$ \operatorname{Cov}(\widehat{\boldsymbol\beta}) \approx \sigma^2(\mathbf X^{\mathsf T}\mathbf W\mathbf X)^{-1}. $$

$\mathbf X^{\mathsf T}\mathbf W\mathbf X$ 的最小特徵值越小，對齊參數越不穩定。因而質量評估除了查看校準誤差，還應檢查重疊面積、源值方差和法方程條件數。

魯棒重加權與局部失敗

遮擋邊界、鏡面高光、錯誤深度和單幀分解失敗會形成離群殘差。把平方損失替換為魯棒 M 估計：

$$ \min_{a,b} \sum_{p\in\Omega_{ir}} \omega_p \rho_{\delta}(a x_p+b-y_p), $$

以 Huber 損失為例 [@huber1964robust]，

$$ \rho_{\delta}(r)= \begin{cases} \frac12r^2, & |r|\leq\delta,\ \delta\left(|r|-\frac12\delta\right), & |r|>\delta. \end{cases} $$

它對應的迭代重加權係數為

$$ u(r)=\frac{\rho_{\delta}'(r)}{r} =\begin{cases} 1, & |r|\leq\delta,\ \delta/|r|, & |r|>\delta. \end{cases} $$

每輪用 $\omega_pu(r_p)$ 更新加權最小二乘，能夠降低孤立大殘差的影響。然而，若一整塊牆面都被預測成錯誤色調，這些殘差可能彼此一致，並不會表現為統計離群點。魯棒估計保護的是參數免受少量異常值支配，不負責判斷模型偏差是否具有空間結構。

可以用有效樣本量觀察權重是否集中在少數像素：

$$ N_{\mathrm{eff}} =\frac{\left(\sum_p\widetilde\omega_p\right)^2} {\sum_p\widetilde\omega_p^2}, \qquad \widetilde\omega_p=\omega_pu(r_p). $$

即使 $|\Omega_{ir}|$ 很大，若大多數像素被降權，$N_{\mathrm{eff}}$ 仍會很小。僅報告重疊像素數量會高估校準的實際支撐。

多視圖圖上的規範傳播

逐幀都對齊到單一參考視角會讓參考幀的局部錯誤傳播到整個場景。把視角看成圖 $\mathcal G=(\mathcal V,\mathcal E)$，邊 $(i,j)$ 上估計相對仿射參數。為便於推導，先考慮無偏置的對數尺度

$$ g_{ij}=\log a_{ij}. $$

每個視角具有未知規範 $s_i=\log a_i$，理想情況下

$$ g_{ij}=s_j-s_i+\epsilon_{ij}. $$

全局估計可以寫成

$$ \min_{{s_i}} \sum_{(i,j)\in\mathcal E} w_{ij}(s_j-s_i-g_{ij})^2, \qquad s_r=0. $$

其法方程由加權圖拉普拉斯矩陣給出。固定一個參考規範後，只有與參考節點連通的視角才能被確定；弱連接節點的不確定性會明顯增大。這個圖結構比“每幀是否存在預測文件”更接近多視圖監督的真實可用性。

環路還能提供不依賴絕對規範的一致性檢查。對閉環 $i\rightarrow j\rightarrow k\rightarrow i$，理想情況下

$$ g_{ij}+g_{jk}+g_{ki}=0. $$

帶偏置的仿射變換記為 $T_{ij}(x)=a_{ij}x+b_{ij}$，則複合閉環應滿足

$$ T_{ki}\circ T_{jk}\circ T_{ij}\approx\operatorname{Id}. $$

較大的閉環誤差說明至少一條邊的對應或校準不可靠。它仍不能定位哪一幀必然錯誤，但能發現單參考誤差指標無法暴露的相互矛盾。

校準一致性與反照率偏差

記校準前後在驗證重疊域上的誤差為

$$ E_{\mathrm{before}} =\frac{1}{Z} \sum_p\omega_p |\widehat{\mathbf A}_i(p)-\widehat{\mathbf A}_r(q_p)|_1, $$

$$ E_{\mathrm{after}} =\frac{1}{Z} \sum_p\omega_p |\mathbf A_i^{*}(p)-\widehat{\mathbf A}_r(q_p)|_1, \qquad Z=\sum_p\omega_p. $$

改善比例可寫為

$$ \rho_i =\frac{E_{\mathrm{before}}-E_{\mathrm{after}}} {E_{\mathrm{before}}+\epsilon}. $$

$\rho_i>0$ 表示低維校準提高了兩幀的一致性，但它不證明二者接近真實反照率。若兩幀都把陰影錯誤地烘焙進材質，校準後仍可高度一致；若參考幀自身錯誤，其他幀向它靠近反而擴大偏差。質量向量至少應包括

$$ \mathbf q_i= \left[ E_{\mathrm{after}}, \rho_i, r_{\mathrm{overlap}}, N_{\mathrm{eff}}, \kappa_i, E_{\mathrm{cycle}}, u_i \right], $$

分別描述殘餘誤差、改善比例、重疊率、有效樣本量、條件數、閉環誤差與預測不確定性。多維質量不能由一個指標完全替代。

質量路由與監督權重

令渲染或解碼得到的反照率為 $\mathbf A_{\theta,i}(p)$，校準後的偽標籤為 $\mathbf A_i^{*}(p)$。帶路由的監督項可以寫成

$$ \mathcal L_A(t) =\frac{ \sum_{i,p} w_A^{(0)} \eta(t) \Gamma(\mathbf q_i) c_i(p) m_i(p) \ell!\left( \mathbf A_{\theta,i}(p), \mathbf A_i^{*}(p) \right) }{ \sum_{i,p} \eta(t)\Gamma(\mathbf q_i)c_i(p)m_i(p)+\epsilon }. $$

$w_A^{(0)}$ 是基礎權重，$\eta(t)$ 是時間調度，$\Gamma(\mathbf q_i)$ 是場景/幀級質量路由，$c_i(p)$ 是像素級連續置信度，$m_i(p)$ 處理無效區域。將這些因子分開，可以區分四類問題：監督總體量綱、當前訓練階段、整幀是否可信、局部像素是否有效。

離散路由可以取

$$ \Gamma(s_i)= \begin{cases} 1, & s_i=\text{通過},\ \gamma, & s_i=\text{警告},\ 0, & s_i=\text{失敗}, \end{cases} \qquad 0<\gamma<1. $$

它便於記錄和復現實驗，但閾值附近會產生跳變。連續形式可以用標準化風險分數 $r_i$：

$$ \Gamma_i =\sigma!\left( \frac{\tau-r_i}{T} \right), $$

其中 $T$ 控制過渡寬度。連續權重更平滑，代價是質量變量的標定必須穩定；若不同場景的分數分佈不一致，同一個 $\tau$ 並不代表相同風險。

質量路由的誤判代價

設偽標籤正確的概率為 $q$。啟用監督時，正確標籤帶來的期望收益為 $B$，錯誤標籤帶來的代價為 $C$；禁用監督的收益記為零。啟用的期望效用為

$$ U_{\mathrm{on}} =qB-(1-q)C. $$

只有當

$$ q>\frac{C}{B+C} $$

時，啟用監督才更合適。如果錯誤偽標籤會通過共享外觀進一步擾動幾何，$C$ 較大，閾值自然應更保守；若反照率分支與幾何完全隔離且具有強魯棒性，$C$ 較小，可以保留更多弱監督。這說明路由閾值取決於下游損害，不只是上游預測指標。

質量狀態還會改變梯度偏差。設真實反照率梯度為 $\mathbf g^{*}$，偽標籤產生的梯度為

$$ \widetilde{\mathbf g} =\mathbf g^{*}+\mathbf b+\boldsymbol\xi, $$

$\mathbf b$ 是系統偏差，$\boldsymbol\xi$ 是零均值噪聲。乘以置信度 $w$ 後，梯度均方誤差為

$$ \mathbb E |w\widetilde{\mathbf g}-\mathbf g^{}|_2^2 =|(w-1)\mathbf g^{}+w\mathbf b|_2^2 +w^2\mathbb E|\boldsymbol\xi|_2^2. $$

降權能壓低噪聲方差，也會削弱有效信號；若偏差方向與真實梯度相反，較大的 $w$ 尤其危險。質量路由需要在信息增益與偏差傳播之間取得可解釋的折中，偽標籤的使用數量並非優先目標。

時間調度與幾何狀態

訓練初期的投影、覆蓋率和外觀特徵都在快速變化。過早啟用跨視角反照率監督，可能把錯誤對應固定到共享表示；過晚啟用則讓顏色參數先吸收照明。一個平滑的升權函數可以寫為

$$ \eta(t) =\begin{cases} 0, & t<t_0,\ \frac12\left[ 1-\cos!\left( \pi\frac{t-t_0}{t_1-t_0} \right) \right], & t_0\leq t<t_1,\ 1, & t\geq t_1. \end{cases} $$

固定迭代數只是近似。更直接的啟用條件可以依賴幾何狀態，例如覆蓋率、深度方差、重投影殘差或對應閉環誤差。當這些量穩定後再提高 $\eta$，調度會更接近監督實際可靠性的變化。

質量也可能隨訓練發生改變。若偽標籤對齊完全基於固定初始幾何，後期幾何改善不會反映到對應；若每輪都在線重算，標籤與模型又形成強反饋迴路。常見折中是低頻更新對應與質量緩存，並用滯回閾值避免狀態在邊界附近反覆切換：

$$ s_i^{(t+1)}= \begin{cases} \text{通過}, & q_i>\tau_{\mathrm{on}},\ \text{失敗}, & q_i<\tau_{\mathrm{off}},\ s_i^{(t)}, & \tau_{\mathrm{off}}\leq q_i\leq\tau_{\mathrm{on}}, \end{cases} $$

其中 $\tau_{\mathrm{off}}<\tau_{\mathrm{on}}$。

校準與路由的適用範圍

多視圖反照率偽監督需要把圖像、對應、校準參數、掩碼和質量狀態作為一個整體保存。缺少其中任何一項，後續訓練都難以區分“沒有標籤”“標籤未對齊”“標籤不可靠”和“標籤只在局部有效”。

校準解決的是規範差異，路由控制的是錯誤傳播。它們都不改變 3DGS 的外觀參數化：即使偽標籤已經足夠穩定，球諧係數、逐高斯潛在特徵和顏色解碼器仍會以不同方式吸收監督並影響幾何。《無球諧外觀表示與神經顏色解碼》從這條梯度路徑繼續分析外觀參數化的影響。

@inproceedings{kocsis2024intrinsic,
  author = {Kocsis, Peter and Sitzmann, Vincent and Nießner, Matthias},
  title = {Intrinsic Image Diffusion for Indoor Single-View Material Estimation},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages = {5198--5208},
  year = {2024}
}

@article{huber1964robust,
  author = {Huber, Peter J.},
  title = {Robust Estimation of a Location Parameter},
  journal = {The Annals of Mathematical Statistics},
  volume = {35},
  number = {1},
  pages = {73--101},
  year = {1964}
}

本徵圖像（Intrinsic Image）提供了一種中間描述：把相對穩定的表面反射屬性與隨照明變化的成分分開。它並不自動給出場景的真實材質，也不能消除單張圖像中的固有歧義；更合適的定位是帶有物理含義、同時受模型假設約束的外觀先驗 [@garces2022survey]。

反射方程與像素乘積

表面點 $\mathbf x$ 沿觀察方向 $\boldsymbol\omega_o$ 的出射輻亮度可寫為

$$ L_o(\mathbf x,\boldsymbol\omega_o) =\int_{\Omega^+} f_r(\mathbf x,\boldsymbol\omega_i,\boldsymbol\omega_o) L_i(\mathbf x,\boldsymbol\omega_i) V(\mathbf x,\boldsymbol\omega_i) (\mathbf n^{\mathsf T}\boldsymbol\omega_i)_+ ,\mathrm d\boldsymbol\omega_i. $$

$f_r$ 是雙向反射分佈函數，$L_i$ 是入射光，$V$ 表示可見性，$\mathbf n$ 是表面法線。若表面近似朗伯反射，則

$$ f_r=\frac{\boldsymbol\rho(\mathbf x)}{\pi}, $$

其中 $\boldsymbol\rho\in[0,1]^3$ 為漫反射反照率。代回積分可得

$$ \mathbf L_o(\mathbf x) =\boldsymbol\rho(\mathbf x)\odot \frac{1}{\pi} \int_{\Omega^+} \mathbf L_iV (\mathbf n^{\mathsf T}\boldsymbol\omega_i)_+ ,\mathrm d\boldsymbol\omega_i. $$

把積分項記為彩色漫反射陰影 $\mathbf S(\mathbf x)$，在線性相機響應、沒有飽和與後處理的近似下，像素形成式為

$$ \mathbf I(p)=\mathbf A(p)\odot\mathbf S(p)+\boldsymbol\varepsilon(p). $$

$\mathbf A$ 是反照率，$\mathbf S$ 同時包含光源強度、光源顏色、法線朝向、遮擋和間接光；$\boldsymbol\varepsilon$ 則吸收鏡面反射、透射、傳感器噪聲以及模型未表示的部分。常見的灰度陰影模型進一步令

$$ \mathbf S(p)=s(p)\mathbf 1, $$

它等價於假設入射光在三個通道上的比例固定。室內混合色溫、彩色牆面二次反射和窗外光進入後，這個條件很容易失效。

分解中的尺度規範

即使乘法模型完全成立，$\mathbf I=\mathbf A\odot\mathbf S$ 仍然沒有唯一解。任取逐通道正數 $\mathbf k$，都有

$$ \mathbf I =(\mathbf k\odot\mathbf A) \odot (\mathbf S\oslash\mathbf k). $$

因此，反照率絕對尺度無法僅由像素乘積恢復。轉到對數域，令

$$ \mathbf i=\log(\mathbf I+\epsilon),\qquad \mathbf a=\log(\mathbf A+\epsilon),\qquad \mathbf s=\log(\mathbf S+\epsilon), $$

則

$$ \mathbf i=\mathbf a+\mathbf s. $$

對任意常量向量 $\boldsymbol\delta$，變換

$$ \mathbf a^{\prime}=\mathbf a+\boldsymbol\delta, \qquad \mathbf s^{\prime}=\mathbf s-\boldsymbol\delta $$

保持觀測不變。其雅可比

$$ \mathbf J= \begin{bmatrix} \mathbf I_3 & \mathbf I_3 \end{bmatrix} $$

至少具有三維零空間

$$ \operatorname{null}(\mathbf J) =\left{ \begin{bmatrix}\boldsymbol\delta\-\boldsymbol\delta\end{bmatrix} :\boldsymbol\delta\in\mathbb R^3 \right}. $$

灰度陰影把自由度從三個通道縮到一個全局尺度，卻仍需要規範條件。固定某一區域的平均陰影、約束反照率動態範圍、使用已知白色表面或在評價時先作尺度對齊，都屬於選擇規範（gauge fixing）；它們讓數值可比較，卻沒有從觀測中創造新的物理信息。

局部變化的分解歧義

對乘法模型作一階微分，有

$$ \delta\mathbf I =\operatorname{diag}(\mathbf S),\delta\mathbf A +\operatorname{diag}(\mathbf A),\delta\mathbf S. $$

若要求像素保持不變，則

$$ \delta\mathbf S =-\operatorname{diag}(\mathbf A)^{-1} \operatorname{diag}(\mathbf S),\delta\mathbf A. $$

任何小的反照率擾動都能被相應的陰影擾動抵消。實際方法必須加入空間或語義先驗，決定哪些變化更像材質、哪些變化更像照明。例如，反照率通常被假設為分段平滑且在材質邊界跳變；陰影通常變化較緩，但投影陰影邊緣又可以非常銳利。簡單的“高頻歸反照率、低頻歸陰影”無法處理牆紙緩變、軟陰影邊界和帶紋理光源。

序關係為尺度歧義之外增加了另一類約束。若人工或模型判斷像素 $p$ 與 $q$ 屬於相同反照率，可約束 $a(p)-a(q)\approx0$；若判斷一個表面更亮，則約束差值符號。Ordinal Shading 通過陰影序關係和全局結構組織分解，說明相對判斷能夠在沒有絕對反照率標定時提供信息 [@careaga2023intrinsic]。不過，序關係只縮小可行集，仍不會給出唯一的輻射尺度。

反照率的條件分佈

單張圖像同時容許多組材質與光照解釋時，確定性模型只能把它們壓到一個輸出。設輸入圖像為 $\mathbf x$，材質屬性圖為 $\mathbf m$，使用均方誤差訓練的確定性預測器滿足

$$ f^*(\mathbf x) =\arg\min_{\mathbf y} \mathbb E!\left[ |\mathbf m-\mathbf y|_2^2 \mid \mathbf x \right] =\mathbb E[\mathbf m\mid\mathbf x]. $$

條件分佈 $p(\mathbf m\mid\mathbf x)$ 若含有多個峰，條件均值可能落在任何一個峰之外。對反照率來說，表現為材質邊界變軟、小結構被平均，或同一區域混合了相互排斥的材質解釋。機率模型改為描述這個條件分佈，在同一 $\mathbf x$ 下通過不同隨機噪聲產生 $\mathbf m^{(1)},\ldots,\mathbf m^{(K)}$。

Intrinsic Image Diffusion 先把反照率與 BRDF 屬性編碼到潛空間。對潛在材質 $\mathbf z_0$ 加噪得到

$$ \mathbf z_t =\sqrt{\bar\alpha_t},\mathbf z_0 +\sqrt{1-\bar\alpha_t},\boldsymbol\epsilon, \qquad \boldsymbol\epsilon\sim\mathcal N(\mathbf 0,\mathbf I), $$

並以輸入圖像的條件特徵 $c(\mathbf x)$ 預測噪聲：

$$ \mathcal L_{\mathrm{diff}} =\mathbb E_{t,\mathbf z_0,\boldsymbol\epsilon} \left[ \left| \boldsymbol\epsilon -\boldsymbol\epsilon_\theta(\mathbf z_t,t,c(\mathbf x)) \right|_2^2 \right]. $$

去噪過程對應從條件分佈中取樣，每個樣本仍需滿足輸入圖像中可觀測的邊界、語義和幾何線索。訓練集來自含有反照率、粗糙度和金屬度標籤的合成場景，預訓練圖像先驗則為真實圖像提供高頻細節與物體語義線索 [@kocsis2024intrinsic]。

彩色照明與殘差分量

真實圖像中，把全部非反照率變化塞入單通道 $s$ 會產生系統誤差。一個更寬的分量模型可以寫成

$$ \mathbf I =\mathbf A_d\odot\mathbf S_d +\mathbf R, $$

$\mathbf A_d$ 和 $\mathbf S_d$ 分別表示漫反射反照率與彩色漫反射陰影，$\mathbf R$ 表示非漫反射殘差。進一步分解

$$ \mathbf R=\mathbf R_+-\mathbf R_-, \qquad \mathbf R_+,\mathbf R_-\geq0, $$

可以分別表示在漫反射預測之上增加亮度的鏡面/發光分量，以及曝光截斷或模型過估導致的負殘差。Colorful Diffuse Intrinsic Image Decomposition 使用彩色陰影與殘差層處理多光源、二次反射和高光，給出了比灰度乘法更接近複雜圖像的描述 [@careaga2024colorful]。

分量增加會降低單個分支的建模偏差，同時擴大解空間。對

$$ \widehat{\mathbf I} =\mathbf A_d\odot\mathbf S_d+\mathbf R $$

而言，僅使用重建損失

$$ \mathcal L_{\mathrm{rec}} =|\widehat{\mathbf I}-\mathbf I|_1 $$

不能阻止 $\mathbf R$ 吸收全部圖像，也不能阻止 $\mathbf A_d$ 與 $\mathbf S_d$ 交換內容。分支必須具有不同的容量、輸入和正則：反照率需要跨照明穩定，陰影應與幾何和光照相關，殘差應保持稀疏或低容量。分解的意義來自這些結構約束，而非變量名稱。

相機管線帶來的額外歧義

相機記錄值通常不是線性輻亮度。把曝光時間記為 $t$、每通道增益記為 $\mathbf g$、相機響應與色調映射記為 $h$，則

$$ \mathbf I_{\mathrm{obs}} =Q!\left( h!\left( \operatorname{clip} (t,\mathbf g\odot\mathbf L,0,L_{\max}) \right) \right)+\boldsymbol\eta, $$

其中 $Q$ 表示量化。若直接在顯示域應用 $\mathbf I=\mathbf A\odot\mathbf S$，響應曲線會改變乘法關係；白平衡會製造逐通道尺度；局部色調映射還會讓相鄰像素的響應彼此相關。

飽和區域的信息損失更直接。標量通道滿足

$$ I_{\mathrm{obs}}=\min(tL,L_{\max}), $$

當 $tL>L_{\max}$ 時，

$$ \frac{\partial I_{\mathrm{obs}}}{\partial L}=0. $$

高光區域中多個不同的輻亮度會映射為同一上限值，分解模型只能依靠鄰域和數據先驗猜測被截斷的內容。暗部量化與讀出噪聲也會降低有效信噪比。因此，反照率置信度應當在飽和、極暗、鏡面和透明區域下降，而不能只根據預測圖是否平滑來判斷。

多視圖條件下的可辨識性

設同一三維表面點在視角 $v$ 中的觀測為

$$ \mathbf I_v =\mathbf A\odot\mathbf S_v+\boldsymbol\varepsilon_v. $$

共享的 $\mathbf A$ 為多視圖提供了耦合。若多個視角的光照調製不同，而反照率保持不變，聯合觀測可以排除一部分單幀解釋。對數域中

$$ \mathbf i_v=\mathbf a+\mathbf s_v, $$

兩視角作差後得到

$$ \mathbf i_v-\mathbf i_u =\mathbf s_v-\mathbf s_u, $$

共享反照率被消去。這有助於估計相對照明變化，卻仍保留共同尺度自由度。更重要的是，只有對應到同一表面點、沒有遮擋切換且材質近似視角無關時，共享假設才成立。位姿誤差、深度誤差和鏡面反射都會讓所謂的跨視角反照率差異混入幾何與視角效應。

若把像素對應寫為由幾何參數 $\boldsymbol\theta_g$ 決定的映射 $q=\pi_u(\mathbf X(p;\boldsymbol\theta_g))$，反照率一致性殘差為

$$ \mathbf r_{uv}(p) =\widehat{\mathbf A}_v(p) -\widehat{\mathbf A}_u(q). $$

其幾何雅可比包含

$$ \frac{\partial\mathbf r_{uv}} {\partial\boldsymbol\theta_g} =- \nabla\widehat{\mathbf A}_u(q) \frac{\partial q}{\partial\boldsymbol\theta_g}. $$

反照率在平坦區域的梯度很小，跨視角一致性對幾何的直接約束仍然會變弱；它主要減少光照變化造成的偽殘差，不能憑空恢復弱紋理區域的深度。這個結論與前文的幾何可觀測性分析一致：更乾淨的外觀信號能夠減少錯誤梯度，但新增幾何秩仍需深度、法線、邊界或多視圖傳播等證據提供。

本徵分解的實驗結果

Intrinsic Image Diffusion 在 InteriorVerse 合成測試集上對每張圖像取樣 $10$ 組材質解。原文同時報告逐樣本指標的均值（Mean metric）、先對 $10$ 張預測圖取均值後的指標（Mean），以及借助真值選出的最佳樣本（Best） [@kocsis2024intrinsic]。

均值圖像的 PSNR 比逐樣本均值高 $1.28,\mathrm{dB}$，說明隨機樣本中有一部分偏差可被平均抵消。最佳樣本又比均值圖像高 $1.01,\mathrm{dB}$，則顯示取樣集合中確實存在更接近真值的解。這一列需要真值才能選擇，它描述候選集的上限，無法代表實際部署時的可獲得誤差。

與 Zhu 等的確定性基線相比，均值圖像的 FID 從 $46.21$ 降至 $25.42$，相對降幅約為 $44.99%$。FID 比較的是預測集與真值集的特徵分佈；數值下降支持整體分佈更接近數據集，對某一張圖像的材質是否正確仍需逐張判斷。

表中的尺度不變指標、分佈指標與圖中的序關係指標分別覆蓋不同性質。一個反照率先驗需要同時檢查三者，並保留樣本間方差、飽和率以及局部邊界誤差。後續跨視角訓練中，這些量會比單一的 PSNR 或 WHDR 更直接地決定偽標籤權重。

反照率先驗的監督範圍

把二維反照率估計接入 3DGS 時，需要區分三種對象：單幀預測給出的候選解釋、跨視角共享的材質狀態、最終 RGB 渲染。它們不應被默認視為同一個真值。

二維模型給出的反照率更適合作為偽標籤，並攜帶尺度、對齊狀態、有效掩碼和置信度。若直接以固定權重監督三維外觀，單幀模型的錯誤會被寫入共享場景表示；若完全捨棄它，又失去複雜光照下有價值的材質線索。《跨視角反照率校準與質量路由》將這個問題轉化為明確的估計與決策過程：在重疊區域校準多視圖反照率，區分可校準漂移與局部失敗，並讓質量狀態控制監督進入訓練的強度。

@article{garces2022survey,
  author = {Garces, Elena and Rodriguez-Pardo, Carlos and Casas, Dan and Lopez-Moreno, Jorge},
  title = {A Survey on Intrinsic Images: Delving Deep Into Lambert and Beyond},
  journal = {International Journal of Computer Vision},
  volume = {130},
  number = {3},
  pages = {836--868},
  year = {2022}
}

@article{careaga2023intrinsic,
  author = {Careaga, Chris and Aksoy, Yağız},
  title = {Intrinsic Image Decomposition via Ordinal Shading},
  journal = {ACM Transactions on Graphics},
  volume = {43},
  number = {1},
  pages = {1--24},
  year = {2023}
}

@inproceedings{kocsis2024intrinsic,
  author = {Kocsis, Peter and Sitzmann, Vincent and Nießner, Matthias},
  title = {Intrinsic Image Diffusion for Indoor Single-View Material Estimation},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages = {5198--5208},
  year = {2024}
}

@article{bell2014intrinsic,
  author = {Bell, Sean and Bala, Kavita and Snavely, Noah},
  title = {Intrinsic Images in the Wild},
  journal = {ACM Transactions on Graphics},
  volume = {33},
  number = {4},
  pages = {159:1--159:12},
  year = {2014}
}

@inproceedings{li2020inverse,
  author = {Li, Zhengqin and Shafiei, Mohammad and Ramamoorthi, Ravi and Sunkavalli, Kalyan and Chandraker, Manmohan},
  title = {Inverse Rendering for Complex Indoor Scenes: Shape, Spatially-Varying Lighting and SVBRDF from a Single Image},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages = {2472--2481},
  year = {2020}
}

@inproceedings{zhu2022inverse,
  author = {Zhu, Jingsen and Luan, Fujun and Huo, Yuchi and Lin, Zihao and Zhong, Zhihua and Xi, Dianbing and Wang, Rui and Bao, Hujun and Zheng, Jiaxiang and Tang, Rui},
  title = {Learning-Based Inverse Rendering of Complex Indoor Scenes with Differentiable Monte Carlo Raytracing},
  booktitle = {SIGGRAPH Asia 2022 Conference Papers},
  pages = {6:1--6:8},
  year = {2022}
}

@article{careaga2024colorful,
  author = {Careaga, Chris and Aksoy, Yağız},
  title = {Colorful Diffuse Intrinsic Image Decomposition in the Wild},
  journal = {ACM Transactions on Graphics},
  volume = {43},
  number = {6},
  pages = {1--12},
  year = {2024}
}

3DGS 的 Alpha 混合並不專屬於 RGB。只要為每個高斯指定一組屬性，同一套可見性權重就能在屏幕上合成相應通道。問題的關鍵隨之轉向三個更細的環節：屬性的像素定義是否符合幾何含義，通道損失的梯度是否真正抵達幾何參數，不同來源的監督是否在當前狀態下彼此相容。

Alpha 混合的通用屬性形式

考慮像素 $p$ 上按深度由近到遠排列的 $N$ 個高斯。第 $i$ 個高斯的有效不透明度、到達該高斯前的透射率及其混合權重分別為 [@kerbl2023gaussians]

$$ \alpha_i(p)=o_iG_i(p), \qquad T_i(p)=\prod_{j<i}\left(1-\alpha_j(p)\right), \qquad w_i(p)=T_i(p)\alpha_i(p). $$

若每個高斯攜帶 $d$ 維屬性 $\mathbf a_i\in\mathbb R^d$，背景屬性為 $\mathbf a_{\mathrm{bg}}$，像素通道可以統一寫成

$$ \mathbf A(p) =\sum_{i=1}^{N}w_i(p)\mathbf a_i +T_{N+1}(p)\mathbf a_{\mathrm{bg}}, $$

其中

$$ T_{N+1}(p)=\prod_{i=1}^{N}\left(1-\alpha_i(p)\right). $$

令 $\mathbf a_i=\mathbf c_i$ 得到 RGB，令 $\mathbf a_i=z_i$ 得到深度矩，令 $\mathbf a_i=\mathbf n_i$ 得到法線向量和；高維的語義或外觀特徵也能以相同方式累積。光柵化器仍然執行一次投影、分塊、排序和前向遍歷，附加通道主要增加每個像素需要維護的累加量。

混合權重之和具有一個有用的閉式形式。由

$$ w_i=T_i-T_{i+1} $$

可得

$$ W(p)=\sum_{i=1}^{N}w_i(p) =1-T_{N+1}(p). $$

$W$ 是前景累積不透明度，也可以看作當前高斯集合對像素的覆蓋程度。當 $W\approx0$ 時，顏色可以由背景補齊，深度和法線卻沒有可靠的前景支撐；因此，幾何通道通常需要保留 $W$，用於歸一化、掩碼或置信度估計。

通道梯度的兩條路徑

設 $\boldsymbol\theta$ 包含全部高斯參數。忽略固定背景後，通用屬性通道的導數為

$$ \frac{\partial\mathbf A(p)}{\partial\boldsymbol\theta} =\sum_i \left[ \mathbf a_i \frac{\partial w_i(p)}{\partial\boldsymbol\theta} +w_i(p) \frac{\partial\mathbf a_i}{\partial\boldsymbol\theta} \right]. $$

第一項經過混合權重傳遞，會改變投影位置、二維 footprint、不透明度與遮擋關係；第二項更新高斯攜帶或由高斯推導的屬性。兩條路徑是否同時存在，取決於通道的定義。

以法線為例，若 $\mathbf n_i$ 是每個高斯獨立學習的向量，法線損失可以直接修改 $\mathbf n_i$，對尺度和旋轉的影響主要經過 $w_i$ 間接產生。若 $\mathbf n_i$ 由協方差的最短軸推導，$\partial\mathbf n_i/\partial\boldsymbol\theta$ 會把法線誤差直接傳給旋轉和尺度。兩種設計輸出同樣是法線圖，對幾何的約束強度卻不同。

混合權重本身也揭示了遮擋下的梯度衰減。對第 $k$ 個不透明度有

$$ \frac{\partial w_i}{\partial\alpha_k}

\begin{cases} 0, & i<k,\ T_k, & i=k,\ -\dfrac{w_i}{1-\alpha_k}, & i>k. \end{cases} $$

前方高斯的 $\alpha_k$ 增大時，自身貢獻增加，後方所有層的權重同時下降。對直接攜帶的屬性，還有

$$ \frac{\partial\mathbf A}{\partial\mathbf a_k}=w_k\mathbf I_d. $$

遠處高斯若被前景遮擋，$T_k$ 和 $w_k$ 都會很小；即使該位置的深度或法線誤差很大，單個視角也難以向它傳遞有效梯度。多通道渲染共享可見性，因此不能繞過遮擋，只能從其他視角、外部觀測或空間先驗補充約束。

深度通道的統計含義

把相機空間中的高斯深度記為 $z_i$。直接替換 RGB 屬性得到的是一階深度矩

$$ M_1(p)=\sum_iw_i(p)z_i, $$

它同時受表面深度和覆蓋率影響。相同平面在圖像邊緣的 $W$ 較低時，$M_1$ 會向零收縮，不能直接解釋為表面深度。常用的期望深度因此寫成 [@huang20242dgs; @turkulainen2025dnsplatter]

$$ D_{\mathrm{mean}}(p) =\frac{M_1(p)}{W(p)+\varepsilon} =\frac{\sum_iw_i(p)z_i} {\sum_iw_i(p)+\varepsilon}. $$

暫時忽略 $\varepsilon$，它對某個權重的導數為

$$ \frac{\partial D_{\mathrm{mean}}}{\partial w_k} =\frac{z_k-D_{\mathrm{mean}}}{W}. $$

位於當前均值前方的高斯接收負梯度，後方高斯接收正梯度，偏離均值越遠，權重變化對期望深度的影響越大。分母中的 $W$ 也說明低覆蓋像素存在放大效應：$W$ 很小時，微小權重擾動就可能令深度劇烈變化。實際監督需要對低 $W$ 區域設置有效掩碼，或者讓置信度隨 $W$ 連續衰減。

二階矩可以進一步描述同一光線上的深度分散程度：

$$ M_2(p)=\sum_iw_i(p)z_i^2, $$

$$ \sigma_z^2(p) =\frac{M_2(p)}{W(p)+\varepsilon} -D_{\mathrm{mean}}^2(p). $$

兩個場景可能具有相近的 $D_{\mathrm{mean}}$，其中一個由窄而集中的權重峰形成，另一個由前後兩層半透明高斯共同平均得到。深度均值看不出這項差異，$\sigma_z^2$ 則會直接暴露沿射線的厚度。

在 $W>0$ 且省略數值穩定項 $\varepsilon$ 的有效像素上，考慮平方形式的兩兩深度失真項

$$ \mathcal L_{\mathrm{dist}}(p) =\sum_{i,j}w_iw_j(z_i-z_j)^2, $$

展開後可以得到

$$ \begin{aligned} \mathcal L_{\mathrm{dist}} &=2WM_2-2M_1^2\ &=2W^2\sigma_z^2. \end{aligned} $$

因此，深度失真正則本質上在壓縮可見權重的深度方差。2DGS 沿用 Mip-NeRF 360 的射線失真思想，使用相近的兩兩距離形式，使沿同一光線分散的圖元向較薄的表面集中 [@barron2022mipnerf360; @huang20242dgs]。這項假設適合單一不透明表面，在玻璃、植被、毛髮或真實多層結構中需要降低權重，否則會把合理的深度多峰分佈強行壓成單峰。

中值深度與可導性

期望深度容易受遠處小權重離群點影響。另一種做法是沿前向合成順序尋找累積不透明度首次越過閾值 $\tau$ 的位置：

$$ k^=\min\left{k:\sum_{i\leq k}w_i\geq\tau\right}, \qquad D_{\mathrm{median}}=z_{k^}. $$

常見的 $\tau=0.5$ 對應近似的中值表面。若 $W<\tau$，該像素沒有越過閾值的有效表面，應當標記為無效或採用明確的回退規則。中值深度對少量遠處浮點更穩健，也更適合把多視圖深度融合為網格；但 $k^*$ 是離散索引，權重尚未跨過閾值時深度不變，跨過閾值後又會突然跳到另一層。它適合幾何提取和診斷，直接作為訓練通道時通常需要平滑近似或額外處理不可導的閾值選擇。

法線通道與局部表面方向

三維高斯的協方差可寫為

$$ \boldsymbol\Sigma_i =\mathbf R_i \operatorname{diag}(s_{i1}^2,s_{i2}^2,s_{i3}^2) \mathbf R_i^{\mathsf T}. $$

當一個尺度明顯小於另外兩個尺度時，高斯接近薄橢球，其最短軸可以近似局部法線。令

$$ m_i=\arg\min_{r\in{1,2,3}}s_{ir}, \qquad \mathbf n_i=\mathbf R_i\mathbf e_{m_i}. $$

$\mathbf n_i$ 的符號需要相對相機或鄰域統一，因為 $\mathbf n$ 和 $-\mathbf n$ 描述同一個無向切平面。最短軸索引在尺度次序交換時也會發生離散跳變；當三個尺度接近時，高斯仍是體積圖元，法線方向本身就缺少穩定含義。

把相機座標系下的法線以可見性權重合成，得到

$$ \widetilde{\mathbf N}(p) =\sum_iw_i(p)\mathbf n_i, \qquad \mathbf N(p) =\frac{\widetilde{\mathbf N}(p)} {|\widetilde{\mathbf N}(p)|_2+\varepsilon}. $$

若忽略 $\varepsilon$ 且令 $r=|\widetilde{\mathbf N}|_2$，歸一化的雅可比為

$$ \frac{\partial\mathbf N} {\partial\widetilde{\mathbf N}} =\frac{1}{r} \left(\mathbf I-\mathbf N\mathbf N^{\mathsf T}\right). $$

$\mathbf I-\mathbf N\mathbf N^{\mathsf T}$ 會去掉平行於當前法線的分量，反向傳播主要調整方向。若前後層法線相反，向量和可能接近零，$1/r$ 會放大數值波動；這也是法線符號、遮擋邊界掩碼和低覆蓋過濾不能省略的原因。

另一條法線來源是渲染深度。對像素齊次座標 $\widetilde{\mathbf p}=(u,v,1)^{\mathsf T}$，反投影點為

$$ \mathbf X(u,v) =D(u,v)\mathbf K^{-1}\widetilde{\mathbf p}. $$

局部切向量給出深度法線

$$ \mathbf N_D(u,v) =\frac{ \partial_u\mathbf X\times\partial_v\mathbf X }{ |\partial_u\mathbf X\times\partial_v\mathbf X|_2+\varepsilon }. $$

渲染法線 $\mathbf N$ 來自高斯的方向與權重，深度法線 $\mathbf N_D$ 來自鄰域深度變化。二者一致時，單個高斯的局部朝向、沿射線的表面位置和圖像鄰域中的幾何變化互相支持。常用一致性項為

$$ \mathcal L_{\mathrm{normal}} =\sum_{p\in\Omega_v} \left(1-\mathbf N(p)^{\mathsf T}\mathbf N_D(p)\right), $$

其中 $\Omega_v$ 應排除低覆蓋、深度不連續和法線退化區域。深度的有限差分會把單像素噪聲傳播到相鄰法線，邊界兩側還可能屬於不同表面；不加選擇地平滑整張法線圖，會令真實折線和薄結構一同變鈍。

正則項與幾何提取的定量影響

2DGS 在 DTU 上使用 Accuracy、Completion 及兩者的平均值評估重建表面，三項均為距離誤差，數值越低越好。消融實驗同時變化了訓練正則、深度統計量和網格提取方法，可以看出最終幾何誤差並非只由渲染模型決定 [@huang20242dgs]。

移除法線一致性後，Average 由 $0.83$ 上升到 $1.24$，相對增幅約為

$$ \frac{1.24-0.83}{0.83}\times100%\approx49.4%. $$

移除深度失真後，Average 增至 $0.88$，相對增幅約為 $6.0%$。在這組 DTU 實驗中，法線一致性對平均誤差的影響更大；這個結果依賴訓練數據的視角覆蓋、場景的局部平面結構和後續網格提取策略，不能直接推廣到透明、多層或極薄表面。

期望深度的 Average 為 $0.94$，差於完整模型使用的中值深度。這與前面的導數分析一致：期望深度對每個權重的敏感度為 $(z_k-D_{\mathrm{mean}})/W$，少量遠離主表面的權重就能拉動均值；中值深度只取累積不透明度跨過閾值的主要表面，用於 TSDF 融合時對小權重浮點更穩健。SPSR 的 Average 為 $1.07$，則說明將高斯中心和法線轉換為網格時，若未顯式利用圖元的不透明度和尺度，渲染模型中已經學到的可見性信息會在提取階段損失。

因此，正則項的幾何收益需要沿完整管線解讀：訓練階段的法線一致性限制局部方向，深度失真壓縮沿射線的權重分佈，深度統計量決定來自多層高斯的信息如何被折疊，網格提取又決定這些信息如何落到最終表面。只看渲染法線或單個平均誤差，都會隱藏這幾個環節之間的誤差傳遞。

特徵通道與解碼位置

高維特徵的合成形式為

$$ \mathbf F(p)=\sum_iw_i(p)\mathbf f_i. $$

它可以用於語義分類、材質預測或顏色解碼。非線性解碼器 $g(\cdot)$ 放在 Alpha 混合前後，會形成兩種不同的模型：

$$ \mathbf C_{\mathrm{decode\text{-}then\text{-}blend}} =\sum_iw_i g(\mathbf f_i), $$

$$ \mathbf C_{\mathrm{blend\text{-}then\text{-}decode}} =g!\left( \frac{\sum_iw_i\mathbf f_i}{W+\varepsilon} \right). $$

只在線性 $g$ 和適當歸一化條件下，兩者才可能等價。前者在高斯層面生成屬性，遮擋關係清晰，但每個高斯都要執行解碼；後者先在像素上匯聚特徵，計算更集中，解碼結果會依賴當前視角下共同覆蓋該像素的高斯集合。模型將材質、光照或視角相關外觀放在哪一側，會改變通道的物理解釋，也會改變輔助損失能夠抵達的參數塊。

這一差別對偽監督尤其重要。若監督目標描述的是表面固有屬性，而解碼器輸入已經混合了多個深度層，遮擋邊界處的標籤誤差可能同時更新前景和背景高斯。屬性通道的名字不能保證其可解釋性，合成位置、歸一化方式和參數依賴關係才決定實際含義。

多通道監督的幾何信息

設幾何參數為 $\mathbf g$，外觀及其他屬性參數為 $\mathbf a$，第 $k$ 個通道的殘差在當前狀態附近線性化為

$$ \mathbf r_k \approx \mathbf r_{k,0} +\mathbf J_{k,g},\delta\mathbf g +\mathbf J_{k,a},\delta\mathbf a. $$

所有通道的加權信息矩陣為

$$ \mathbf H =\sum_k \lambda_k \begin{bmatrix} \mathbf J_{k,g}^{\mathsf T}\mathbf W_k\mathbf J_{k,g} & \mathbf J_{k,g}^{\mathsf T}\mathbf W_k\mathbf J_{k,a}\ \mathbf J_{k,a}^{\mathsf T}\mathbf W_k\mathbf J_{k,g} & \mathbf J_{k,a}^{\mathsf T}\mathbf W_k\mathbf J_{k,a} \end{bmatrix}. $$

消去屬性增量後，幾何實際獲得的約化信息仍由 Schur 補決定：

$$ \mathbf H_{g}^{\mathrm{eff}} =\mathbf H_{gg} -\mathbf H_{ga}\mathbf H_{aa}^{-1}\mathbf H_{ag}. $$

新增通道若只重複 RGB 已有的雅可比方向，或者其誤差能被獨立屬性參數完全吸收，$\mathbf H_g^{\mathrm{eff}}$ 的最小特徵值未必明顯提高。深度傳感器、多視圖匹配和可靠法線先驗能夠提供與顏色不同的方向；單目預測、平滑項和深度失真項則通過模型假設壓縮解空間。二者都能改善數值可解性，信息來源需要分開表述：前者引入額外觀測，後者引入外部先驗或結構偏置。

這個區分還能解釋一種常見現象。將可學習法線直接擬合偽標籤時，法線圖可能很快變得平滑，位置和協方差仍然噪雜；從協方差最短軸推導法線時，同樣的誤差會更直接地限制高斯方向，卻依賴“高斯應當趨於扁平”這一額外假設。監督目標、通道參數化和幾何收益之間沒有一一對應關係。

監督衝突與權重

把各通道損失對共享參數的梯度記為

$$ \mathbf g_k=\nabla_{\boldsymbol\theta}\mathcal L_k, $$

總更新為

$$ \Delta\boldsymbol\theta =-\eta\sum_j\lambda_j\mathbf g_j. $$

對第 $k$ 個損失作一階展開：

$$ \begin{aligned} \Delta\mathcal L_k &\approx \mathbf g_k^{\mathsf T}\Delta\boldsymbol\theta\ &=-\eta\lambda_k|\mathbf g_k|2^2 -\eta\sum{j\ne k} \lambda_j\mathbf g_k^{\mathsf T}\mathbf g_j. \end{aligned} $$

第一項保證該損失沿自身負梯度下降，第二項描述其他通道帶來的幫助或衝突。可用梯度夾角

$$ \cos\phi_{kj} =\frac{ \mathbf g_k^{\mathsf T}\mathbf g_j }{ |\mathbf g_k|_2|\mathbf g_j|_2 } $$

觀察局部關係。$\cos\phi_{kj}>0$ 時，兩項在當前附近相互促進；$\cos\phi_{kj}<0$ 時，一項的下降方向會推高另一項。遮擋邊界處，RGB 可能希望移動高斯以修正顏色輪廓，平滑法線項卻傾向於保持鄰域方向一致；錯誤的單目深度也可能與多視圖光度證據直接衝突。

損失數值的量綱同樣會影響更新。米制深度誤差、單位法線的角度誤差和 $[0,1]$ RGB 殘差不能僅按標量大小比較。除了固定權重，還可以把空間有效性、先驗質量和時間調度分開寫成

$$ \mathcal L(t) =\sum_k\sum_{p\in\Omega} \overline\lambda_k s_k(t) q_k(p) m_k(p) \ell_k(p). $$

$\overline\lambda_k$ 控制通道的基礎尺度，$s_k(t)$ 描述訓練階段，$q_k(p)$ 表示連續置信度，$m_k(p)$ 則處理遮擋、缺失值和無效區域。訓練初期幾何和可見性尚未穩定，強輔助監督可能固定錯誤對應；後期若始終保持過強平滑，又可能抹去已經形成的細節。時間權重需要針對監督可靠性變化設置，不能只把全部通道同時打開。

若某個觀測殘差 $r(p)$ 的噪聲近似為方差 $\sigma^2(p)$ 的高斯分佈，其負對數似然給出

$$ \ell(p) =\frac{r^2(p)}{2\sigma^2(p)} +\frac12\log\sigma^2(p). $$

$1/\sigma^2$ 自然充當置信度權重，$\log\sigma^2$ 防止通過無限增大不確定性逃避監督。傳感器誤差模型、單目預測置信度、多視圖一致性和渲染覆蓋率都可以進入 $\sigma^2(p)$，比單一全局係數更接近各通道真實的可靠範圍。

多通道監督的適用範圍

多通道輸出既服務訓練，也服務診斷。兩種用途需要保留不同的信息：訓練通道強調可導和穩定，診斷通道還要揭示被歸一化或聚合隱藏的狀態。

只保存歸一化深度會丟失 $W$，只保存法線顏色圖會丟失歸一化前的向量範數，只觀察總損失會混合各通道的量綱和啟用狀態。更完整的訓練記錄應至少保留覆蓋率、深度一二階矩、法線有效掩碼、各子項的未加權值與加權貢獻。這樣才能判斷某次指標改善來自新增幾何約束、外觀參數補償，還是監督區域本身發生了變化。

結語

多通道可微渲染建立在同一組 Alpha 權重上。顏色、深度、法線和特徵因而觀察同一個高斯集合，也共同受到投影誤差、近似排序、遮擋和覆蓋不足的影響。期望深度需要與覆蓋率和深度方差一起解釋，法線合成需要處理方向符號、歸一化和深度邊界，特徵通道還受到非線性解碼位置的影響。

輔助監督能否改善幾何，最終取決於它對幾何參數的雅可比是否提供了新的有效方向。外部深度和多視圖幾何可以增加觀測證據，單目預測、法線平滑與深度失真則引入先驗假設；兩類信息都要經過可見性權重和通道參數化才能抵達高斯。《複雜光照下的本徵圖像與反照率先驗》繼續拆開顏色通道內部的反照率、光照和視角相關外觀，並分析這些分量各自能夠吸收哪些誤差。

@article{kerbl2023gaussians,
  author = {Kerbl, Bernhard and Kopanas, Georgios and Leimkühler, Thomas and Drettakis, George},
  title = {3D Gaussian Splatting for Real-Time Radiance Field Rendering},
  journal = {ACM Transactions on Graphics},
  volume = {42},
  number = {4},
  pages = {139:1--139:14},
  year = {2023}
}

@inproceedings{huang20242dgs,
  author = {Huang, Binbin and Yu, Zehao and Chen, Anpei and Geiger, Andreas and Gao, Shenghua},
  title = {2D Gaussian Splatting for Geometrically Accurate Radiance Fields},
  booktitle = {ACM SIGGRAPH 2024 Conference Papers},
  year = {2024}
}

@inproceedings{turkulainen2025dnsplatter,
  author = {Turkulainen, Matias and Ren, Xuqian and Melekhov, Iaroslav and Seiskari, Otto and Rahtu, Esa and Kannala, Juho},
  title = {DN-Splatter: Depth and Normal Priors for Gaussian Splatting and Meshing},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2025}
}

@inproceedings{barron2022mipnerf360,
  author = {Barron, Jonathan T. and Mildenhall, Ben and Verbin, Dor and Srinivasan, Pratul P. and Hedman, Peter},
  title = {Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages = {5470--5479},
  year = {2022}
}

設場景參數為 $\boldsymbol\theta$，觀測模型為 $\mathbf h(\boldsymbol\theta)$，圖像數據為 $\mathbf y$。在當前估計 $\boldsymbol\theta_0$ 附近，殘差可以線性化為

$$ \mathbf r(\boldsymbol\theta_0+\delta\boldsymbol\theta) \approx \mathbf r_0+\mathbf J\delta\boldsymbol\theta, \qquad \mathbf J= \left.\frac{\partial\mathbf r}{\partial\boldsymbol\theta}\right|_{\boldsymbol\theta_0}. $$

若存在非零方向 $\mathbf v$ 滿足 $\mathbf J\mathbf v\approx\mathbf 0$，參數沿 $\mathbf v$ 改變後，當前觀測幾乎不變。這個方向在局部缺少可觀測性。加權最小二乘對應的信息矩陣為

$$ \mathbf H=\mathbf J^{\mathsf T}\mathbf W\mathbf J. $$

$\mathbf H$ 的小特徵值標記弱約束方向，零特徵值則對應局部零空間。弱紋理的影響由此可以拆成三層：圖像是否提供位移方向，視角是否提供足夠視差，以及渲染模型中的幾何變化能否同外觀變化區分開來。

光度殘差的幾何雅可比

取參考圖像 $I_r$ 中的像素 $\mathbf p$，根據深度、相機位姿和投影模型，把它映射到目標圖像中的 $\mathbf w(\mathbf p;\boldsymbol\theta)$。直接法使用的光度殘差可寫為 [@engel2018dso]

$$ r_{\mathbf p}(\boldsymbol\theta) =I_t!\left(\mathbf w(\mathbf p;\boldsymbol\theta)\right)-I_r(\mathbf p). $$

對幾何參數求導，鏈式法則得到

$$ \mathbf J_{\mathbf p}

\underbrace{\nabla I_t(\mathbf w)^{\mathsf T}}{\text{圖像方向}} \underbrace{\frac{\partial\mathbf w}{\partial\boldsymbol\theta}}{\text{投影與幾何}}. $$

第二項描述深度或位姿變化會把像素移到哪裏，第一項決定這次移動能否改變觀測值。當 $|\nabla I_t|$ 很小時，即使投影位置對幾何非常敏感，乘積仍然接近零；圖像沒有為這次幾何變化提供足夠的判別信號。

把圖像噪聲近似為方差 $\sigma_I^2$ 的獨立高斯噪聲，局部參數協方差可由高斯—牛頓近似寫成

$$ \operatorname{Cov}(\delta\boldsymbol\theta) \approx\sigma_I^2(\mathbf J^{\mathsf T}\mathbf J)^{-1}. $$

這個表達式需要 $\mathbf J$ 滿列秩；退化時應使用偽逆，零空間方向的方差沒有有限上界。低 RGB 殘差只表示某個解能解釋訓練圖像，沒有給出 $\mathbf J^{\mathsf T}\mathbf J$ 的譜，也就沒有給出幾何置信度。

局部窗口的秩

先把幾何變化簡化成圖像窗口的二維平移 $\boldsymbol\delta=(u,v)^{\mathsf T}$。亮度恆常假設給出

$$ I(x+u,y+v)\approx I(x,y)+I_xu+I_yv. $$

在窗口 $\Omega$ 中疊加殘差，得到經典的二階矩陣 [@lucas1981iterative; @shi1994good]

$$ \mathbf G =\sum_{\mathbf p\in\Omega}w_{\mathbf p} \begin{bmatrix} I_x^2&I_xI_y\ I_xI_y&I_y^2 \end{bmatrix} =\sum_{\mathbf p\in\Omega}w_{\mathbf p} \nabla I_{\mathbf p}\nabla I_{\mathbf p}^{\mathsf T}. $$

$\mathbf G$ 正是位移參數的信息矩陣。令其特徵值滿足 $\lambda_1\geq\lambda_2\geq0$：

• 平坦窗口中，$\lambda_1$、$\lambda_2$ 都很小，兩個平移方向都缺少約束；
• 單一邊緣通常只有 $\lambda_1$ 較大，沿邊緣方向仍可滑動，形成孔徑問題；
• 角點或方向豐富的紋理使兩個特徵值都較大，二維位移才能穩定估計。

如果每個像素殘差的噪聲方差仍為 $\sigma_I^2$，位移估計的協方差近似為

$$ \operatorname{Cov}(\hat{\boldsymbol\delta}) \approx\sigma_I^2\mathbf G^{-1} =\mathbf V \begin{bmatrix} \sigma_I^2/\lambda_1&0\ 0&\sigma_I^2/\lambda_2 \end{bmatrix} \mathbf V^{\mathsf T}. $$

因而 $\lambda_2$ 直接控制最差方向的方差上界，條件數

$$ \kappa(\mathbf G)=\frac{\lambda_1}{\lambda_2} $$

則描述兩個方向的敏感度差異。只使用平均梯度幅值會把「兩個方向都弱」和「一個方向很強、另一個方向近乎為零」混在一起。對幾何可觀測性而言，最小特徵值和條件數比單個梯度閾值更有解釋力。

視差的深度靈敏度

局部窗口可被穩定匹配後，深度仍取決於視角幾何。在校正後的雙目模型中，焦距為 $f$，基線為 $B$，深度為 $Z$，視差 $d$ 滿足

$$ d=\frac{fB}{Z}, \qquad Z=\frac{fB}{d}. $$

對視差做一階誤差傳播，可得

$$ \sigma_Z \approx\left|\frac{\partial Z}{\partial d}\right|\sigma_d =\frac{Z^2}{fB}\sigma_d, \qquad \frac{\sigma_Z}{Z}\approx\frac{Z}{fB}\sigma_d. $$

深度越遠、基線越小，固定的亞像素匹配誤差會被放大得越嚴重。視差誤差本身又受紋理約束。對一維水平匹配，信息量近似為

$$ \mathcal I_d =\frac{1}{\sigma_I^2}\sum_{\mathbf p\in\Omega}I_x(\mathbf p)^2, \qquad \sigma_d^2 \approx\frac{\sigma_I^2}{\sum_{\mathbf p\in\Omega}I_x(\mathbf p)^2}. $$

把兩段誤差傳播連起來，得到

$$ \sigma_Z^2 \approx \frac{Z^4}{f^2B^2} \frac{\sigma_I^2}{\sum_{\mathbf p\in\Omega}I_x(\mathbf p)^2}. $$

這個式子把弱紋理和視角退化放在了同一個尺度上：低梯度減小分母，遠距離和小基線增大分子。增加視圖數只有在新視圖帶來有效梯度、共視關係和新的射線方向時，才會真正增加信息量。

一般相機模型下可以用兩條觀測射線的夾角 $\theta$ 代替校正雙目的水平視差。當 $\theta\to0$ 時，兩條射線接近平行，三角化信息矩陣沿深度方向的特徵值趨近於零。基線—深度比、共視數量和射線夾角應一起考察，單獨統計「看見了幾次」仍可能高估幾何約束 [@hartley2004multiple]。

位姿與結構的聯合零空間

SfM 在多幅圖像中建立特徵軌跡，並通過光束法平差聯合優化相機與三維點 [@schonberger2016structure; @triggs2000bundle]。第 $i$ 個相機觀測第 $j$ 個點時，重投影殘差可寫成

$$ \mathbf r_{ij} =\mathbf z_{ij}-\pi(\mathbf T_i\mathbf X_j). $$

線性化後，每條觀測只連接一個相機塊和一個點塊：

$$ \delta\mathbf r_{ij} \approx \mathbf J_{c,ij}\delta\boldsymbol\xi_i +\mathbf J_{x,ij}\delta\mathbf X_j. $$

全部觀測形成的法方程具有分塊結構

$$ \begin{bmatrix} \mathbf H_{cc}&\mathbf H_{cx}\ \mathbf H_{xc}&\mathbf H_{xx} \end{bmatrix} \begin{bmatrix} \delta\mathbf c\ \delta\mathbf x \end{bmatrix} =- \begin{bmatrix} \mathbf g_c\ \mathbf g_x \end{bmatrix}. $$

消去三維點後，相機位姿接收到的有效信息由 Schur 補給出

$$ \mathbf S_c =\mathbf H_{cc} -\mathbf H_{cx}\mathbf H_{xx}^{-1}\mathbf H_{xc}. $$

某個三維點只有短軌跡、小三角化角或集中在單一方向的觀測時，$\mathbf H_{xx}$ 會變得病態；這部分不確定性經交叉塊傳到 $\mathbf S_c$，令位姿和結構沿某些組合方向一起漂移。弱紋理在特徵法中常表現為軌跡數量不足，在直接法中則表現為光度雅可比過小，最後都會反映到法方程的譜上。

單目重建還天然具有全局相似變換自由度：同時旋轉、平移和縮放全部相機與三維點，重投影不變。未固定坐標系時，這對應 $3+3+1=7$ 個規範自由度。它和弱紋理造成的局部退化需要分開處理：前者可通過固定參考相機與尺度消除，後者即使在規範固定後仍然存在。

3DGS 的外觀—幾何耦合

3DGS 通常從 SfM 稀疏點初始化高斯位置，然後依靠渲染損失更新位置、尺度、旋轉、不透明度與顏色，並根據屏幕空間位置梯度進行致密化 [@kerbl2023gaussians]。SfM 在弱紋理表面留下的稀疏或不穩定初始化，會直接縮小後續幾何優化的有效支撐域。

對某個像素，按深度排序的 Alpha 混合為

$$ \hat{\mathbf C}(\mathbf p) =\sum_iT_i(\mathbf p)\alpha_i(\mathbf p)\mathbf c_i, \qquad T_i=\prod_{j<i}(1-\alpha_j). $$

令第 $k$ 個高斯的二維核為

$$ G_k(\mathbf p) =\exp!\left[-\frac12 (\mathbf p-\boldsymbol\mu_k')^{\mathsf T} \mathbf Q_k (\mathbf p-\boldsymbol\mu_k')\right], \qquad \alpha_k=o_kG_k, $$

其中 $\mathbf Q_k=(\boldsymbol\Sigma_k')^{-1}$。把第 $k$ 層之後的合成顏色記為 $\mathbf B_{k+1}$，則

$$ \frac{\partial\hat{\mathbf C}}{\partial\alpha_k} =T_k(\mathbf c_k-\mathbf B_{k+1}), $$

$$ \frac{\partial\alpha_k}{\partial\boldsymbol\mu_k'} =o_kG_k\mathbf Q_k(\mathbf p-\boldsymbol\mu_k'). $$

再乘上投影雅可比，就得到世界空間高斯中心的像素梯度。平坦區域中，相鄰位置產生的顏色差異很小，footprint 兩側的殘差也容易對稱抵消；位置梯度隨之變弱。此時致密化缺少觸發信號，增加訓練迭代並不會憑空增加新的幾何證據。

參數耦合會進一步削弱幾何的可觀測性。把位置、旋轉和尺度合記為幾何參數 $\mathbf g$，把顏色與不透明度合記為外觀參數 $\mathbf a$，局部渲染變化為

$$ \delta\hat{\mathbf C} \approx \mathbf J_g\delta\mathbf g+\mathbf J_a\delta\mathbf a. $$

若先允許外觀參數自由調整，再考察幾何剩餘的信息量，可消去 $\delta\mathbf a$，得到幾何的約化信息矩陣

$$ \mathbf S_g =\mathbf J_g^{\mathsf T}\mathbf W\mathbf J_g -\mathbf J_g^{\mathsf T}\mathbf W\mathbf J_a (\mathbf J_a^{\mathsf T}\mathbf W\mathbf J_a)^{-1} \mathbf J_a^{\mathsf T}\mathbf W\mathbf J_g. $$

第一項是固定外觀時的幾何信息，第二項扣除了可以由外觀參數解釋的分量。若 $\mathbf J_g$ 的列接近 $\mathbf J_a$ 的列空間，$\mathbf S_g$ 便會出現小特徵值：移動高斯造成的圖像變化，可以通過調整顏色、不透明度或 footprint 近似抵消。訓練視角的 RGB 因而能繼續改善，表面位置、法線與新視角遮擋關係仍可能不穩定。

傳播與平面約束的實驗分解

GaussianPro 將相鄰高斯的幾何線索傳播到低紋理區域，並用多視圖單應代價篩選深度—法線候選。其 Waymo 消融實驗把幾何傳播和平面約束分別開關，因而可以觀察兩條路徑對渲染結果的單獨與聯合影響 [@cheng2024gaussianpro]。

幾何傳播單獨使 PSNR 由 $33.53$ dB 增至 $34.48$ dB，增量為 $0.95$ dB；只啟用平面約束時，增量為 $0.49$ dB。當傳播已經啟用，再加入平面約束只增加 $0.20$ dB。若兩項影響可以線性相加，完整設置的預期增量應為 $0.95+0.49=1.44$ dB；實際增量為 $1.15$ dB，少了 $0.29$ dB。這部分差值反映出兩項約束會更新重疊的高斯參數，並沿共享的誤差路徑改變相同的渲染結果；後加入的約束只能繼續修正尚未被前一項消除的分量。

表中三項都是圖像指標，只能間接支撐幾何改善。PSNR 和 SSIM 的增長可能同時來自更正確的表面位置、更密集的高斯覆蓋以及外觀參數調整。因此，定量表格要同時對照高斯分佈和法線圖：前者顯示幾何質量是否集中在表面附近，後者顯示局部方向是否連續。這比單獨用 RGB 指標推斷可觀測性更符合前面的約化信息矩陣分析。

弱紋理的幾種退化型態

「弱紋理」常被簡化成梯度小於某個閾值，實際退化還包含方向、對應唯一性、成像假設和視角幾何。它們在信息矩陣中留下的症狀不同。

重複紋理尤其容易被單純的局部特徵值判斷漏掉。每個候選位置附近都可能具有很強的二階曲率，但全局目標包含多個相似極小值；這屬於離散歧義。局部可觀測性回答「當前解附近能否辨識」，還需要多視圖循環一致性、極線約束或更大範圍的上下文排除錯誤分支。

可觀測性的診斷尺度

弱紋理區域的判定可以沿着優化管線逐層記錄，而不必壓縮成單一掩碼：

• 圖像層：統計窗口結構張量的 $\lambda_2$、$\kappa(\mathbf G)$ 和梯度方向分佈；
• 匹配層：保留代價曲線的峰值比、次優候選間隔和前後向一致性；
• 視角層：統計有效共視數、基線—深度比、三角化角與遮擋狀態；
• 優化層：觀察重投影雅可比或近似 Hessian 的小特徵值，以及阻尼前後的條件數；
• 3DGS 層：比較幾何梯度與外觀梯度範數，檢查屏幕空間位置梯度、可見視圖數、深度和法線的跨視圖一致性；
• 結果層：把 RGB 指標同深度、法線、點雲或網格指標分開報告，並保留新視角誤差圖。

正則項與先驗會抬高某些弱方向的曲率。若原始信息矩陣為 $\mathbf H$，加入二次先驗

$$ \frac{\lambda}{2} |\mathbf L(\boldsymbol\theta-\boldsymbol\theta_0)|_2^2 $$

後，法方程變為

$$ \mathbf H_{\mathrm{reg}} =\mathbf H+\lambda\mathbf L^{\mathsf T}\mathbf L. $$

數值條件會改善，但新增曲率來自先驗假設。當數據項在某個方向沒有信息時，結果主要由 $\boldsymbol\theta_0$、$\mathbf L$ 和 $\lambda$ 決定；先驗的偏差也會直接進入幾何。診斷時應把數據可觀測性和正則化後的可解性分開記錄。

結語

弱紋理場景的幾何困難可以沿一條連續的誤差鏈理解：圖像梯度決定局部位移信息，視差和射線夾角把像素不確定性傳到深度，位姿—結構耦合把局部退化擴散到整個重建，而 3DGS 的顏色、不透明度和協方差又能吸收一部分幾何引起的圖像變化。

局部結構張量的最小特徵值、三角化角、光束法平差的 Schur 補和 3DGS 的約化幾何信息矩陣，描述的是同一件事在不同階段的投影：幾何擾動是否會留下無法被其他參數消除的觀測差異。這也解釋了弱紋理平面上常見的現象——訓練視角已經平滑、乾淨，深度、法線和新視角邊界仍然漂移。後續幾何約束應當針對這些弱方向補充跨視圖證據，同時明確新增信息來自觀測、模型假設還是外部先驗。

@inproceedings{lucas1981iterative,
  author = {Lucas, Bruce D. and Kanade, Takeo},
  title = {An Iterative Image Registration Technique with an Application to Stereo Vision},
  booktitle = {Proceedings of the International Joint Conference on Artificial Intelligence},
  pages = {674--679},
  year = {1981}
}

@inproceedings{shi1994good,
  author = {Shi, Jianbo and Tomasi, Carlo},
  title = {Good Features to Track},
  booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages = {593--600},
  year = {1994}
}

@book{hartley2004multiple,
  author = {Hartley, Richard and Zisserman, Andrew},
  title = {Multiple View Geometry in Computer Vision},
  edition = {2},
  publisher = {Cambridge University Press},
  year = {2004}
}

@incollection{triggs2000bundle,
  author = {Triggs, Bill and McLauchlan, Philip F. and Hartley, Richard I. and Fitzgibbon, Andrew W.},
  title = {Bundle Adjustment---A Modern Synthesis},
  booktitle = {Vision Algorithms: Theory and Practice},
  volume = {1883},
  pages = {298--372},
  publisher = {Springer},
  year = {2000}
}

@article{engel2018dso,
  author = {Engel, Jakob and Koltun, Vladlen and Cremers, Daniel},
  title = {Direct Sparse Odometry},
  journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
  volume = {40},
  number = {3},
  pages = {611--625},
  year = {2018}
}

@inproceedings{schonberger2016structure,
  author = {Schönberger, Johannes L. and Frahm, Jan-Michael},
  title = {Structure-from-Motion Revisited},
  booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages = {4104--4113},
  year = {2016}
}

@article{kerbl2023gaussians,
  author = {Kerbl, Bernhard and Kopanas, Georgios and Leimkühler, Thomas and Drettakis, George},
  title = {3D Gaussian Splatting for Real-Time Radiance Field Rendering},
  journal = {ACM Transactions on Graphics},
  volume = {42},
  number = {4},
  pages = {139:1--139:14},
  year = {2023}
}

@inproceedings{cheng2024gaussianpro,
  author = {Cheng, Kai and Long, Xiaoxiao and Yang, Kaizhi and Yao, Yao and Yin, Wei and Ma, Yuexin and Wang, Wenping and Chen, Xuejin},
  title = {GaussianPro: 3D Gaussian Splatting with Progressive Propagation},
  booktitle = {Proceedings of the International Conference on Machine Learning},
  pages = {8123--8140},
  year = {2024}
}

《3DGS 的表示、渲染與優化》將場景表示、投影、Alpha 混合和密度控制串成了完整管線。渲染器給定相機 $\mathcal{C}_m$ 和高斯集合 $\Theta$，輸出訓練視角的渲染圖像

$$ \hat{\mathbf I}_m =\mathcal R(\Theta;\mathcal C_m). $$

訓練還需要把渲染圖像與真實圖像 $\mathbf I_m$ 的差異壓縮成標量損失，再沿可微光柵化器反向傳播。若 $\theta_i$ 表示第 $i$ 個高斯的某個參數，鏈式法則給出

$$ \frac{\partial\mathcal L}{\partial\theta_i} =\sum_{p\in\Omega_i} \frac{\partial\mathcal L}{\partial\hat{\mathbf I}(p)} \frac{\partial\hat{\mathbf I}(p)}{\partial\theta_i}, $$

其中 $\Omega_i$ 是該高斯在當前視角中實際影響的像素集合。右側第一項由損失函數決定，描述每個像素應當向哪個方向修正；第二項由投影、二維高斯權重、Alpha 混合與球諧顏色共同決定，負責把像素修正轉換成位置、尺度、旋轉、不透明度和顏色的梯度。

這條梯度鏈也劃分了兩種不同角色：訓練目標參與參數更新，重建指標在留出視角上描述最終結果。二者可以共享相似的數學形式，使用時的統計範圍和解釋方式卻不同。

flowchart TD
    A["多視角圖像與相機"] --> B["3DGS 渲染器"]
    B --> C{"當前視角用途"}
    C -->|"訓練視角"| D["L1 與 D-SSIM"]
    D --> E["反向傳播"]
    E --> F["更新高斯參數與密度"]
    F -.->|"下一次迭代"| B
    C -->|"留出測試視角"| G["固定模型渲染"]
    G --> H["SSIM / PSNR / L1 / LPIPS"]
    H --> I["指標、曲線與同視角圖像"]

L1：逐像素殘差的直接尺度

把顏色通道和像素位置統一編號為 $j=1,\ldots,N$，令

$$ e_j=\hat I_j-I_j, $$

平均 L1 損失為

$$ \mathcal L_1 =\frac{1}{N}\sum_{j=1}^{N}|e_j|. $$

除 $e_j=0$ 外，它對渲染像素的導數只有符號差異：

$$ \frac{\partial\mathcal L_1}{\partial\hat I_j} =\frac{1}{N}\operatorname{sign}(e_j). $$

因此，大殘差像素不會像 L2 那樣獲得與誤差幅值成正比的更大梯度。少量高亮、遮擋錯配或曝光異常像素對整體更新的支配作用相對較弱，這也是圖像重建中常用 L1 的原因之一。

L1 的穩健性還可以從統計估計的角度理解。把某個待預測的標量記為 $a$，相應觀測記為隨機變量 $Y$，則期望 L1 風險為

$$ R_1(a)=\mathbb E\left[|Y-a|\right]. $$

在分佈連續處，其導數可以寫成

$$ \frac{\mathrm dR_1}{\mathrm da} =\Pr(Y<a)-\Pr(Y>a). $$

令導數為零，得到 $\Pr(Y<a)=\Pr(Y>a)$，所以 L1 對應條件分佈的中位數。作為對照，平方損失滿足

$$ R_2(a)=\mathbb E\left[(Y-a)^2\right], \qquad \frac{\mathrm dR_2}{\mathrm da} =2\left(a-\mathbb E[Y]\right), $$

其最優解是均值。當殘差分佈帶有長尾時，少量極端觀測會直接拉動均值，對中位數的影響則小得多。映射到多視角重建中，遮擋切換、鏡面高光和相機標定殘差都可能形成長尾誤差；L1 由此提供了較穩定的基礎修正信號。

L1 對每個坐標獨立計算。兩片紋理即使具有相似的局部結構，只要發生一兩個像素的位移，逐點殘差仍會迅速上升；反過來，一塊平滑區域只要平均顏色接近，也可能得到很低的 L1，而其邊界位置、局部對比度或細微紋理仍然不準確。3DGS 在 L1 之外加入結構項，補充的正是這部分局部關係。

SSIM：局部亮度、對比度與相關性

結構相似性指標（Structural Similarity Index Measure，SSIM）在局部窗口中比較兩幅圖像的均值、方差和協方差。對窗口內的兩組像素 $x$、$y$，常用形式為 [@wang2004ssim]

$$ \operatorname{SSIM}(x,y) =\frac{(2\mu_x\mu_y+C_1)(2\sigma_{xy}+C_2)} {(\mu_x^2+\mu_y^2+C_1)(\sigma_x^2+\sigma_y^2+C_2)}. $$

若窗口權重 $w_q$ 已歸一化，局部統計量可以寫成

$$ \mu_x=\sum_q w_qx_q, \qquad \sigma_x^2=\sum_qw_q(x_q-\mu_x)^2, \qquad \sigma_{xy}=\sum_qw_q(x_q-\mu_x)(y_q-\mu_y). $$

第一個分式主要比較局部亮度，第二個分式同時反映對比度與結構相關性。常數 $C_1$、$C_2$ 用於穩定低亮度或低方差窗口中的除法；它們通常按像素動態範圍 $L$ 設為 $(K_1L)^2$ 和 $(K_2L)^2$。當圖像已歸一化到 $[0,1]$ 時，常見設置為 $K_1=0.01$、$K_2=0.03$。

整幅圖像的 SSIM 一般由各通道、各窗口中心的局部分數取平均。這個平均看似簡單，具體實現仍包含窗口尺寸、高斯核標準差、邊界填充、通道聚合和動態範圍等選擇。兩組實驗若使用不同設置，最後幾位小數未必具有可比性。

局部統計量也能直接給出 SSIM 的梯度結構。把渲染窗口記為 $x$、真實窗口記為 $y$，並令

$$ \begin{aligned} A&=2\mu_x\mu_y+C_1, &B&=2\sigma_{xy}+C_2,\ C&=\mu_x^2+\mu_y^2+C_1, &D&=\sigma_x^2+\sigma_y^2+C_2, \end{aligned} $$

則 $s=\operatorname{SSIM}(x,y)=AB/(CD)$。對窗口中的第 $r$ 個渲染像素，有

$$ \begin{aligned} \frac{\partial A}{\partial x_r}&=2w_r\mu_y, &\frac{\partial B}{\partial x_r}&=2w_r(y_r-\mu_y),\ \frac{\partial C}{\partial x_r}&=2w_r\mu_x, &\frac{\partial D}{\partial x_r}&=2w_r(x_r-\mu_x). \end{aligned} $$

代入乘法與商法則，得到

$$ \frac{\partial s}{\partial x_r} =\frac{A_r'B+AB_r'}{CD} -\frac{AB(C_r'D+CD_r')}{(CD)^2}. $$

這個導數同時含有目標像素 $y_r$、窗口均值和局部離均差。即使只對 $x_r$ 求導，修正方向仍由整個窗口的統計狀態決定；窗口重疊後，一個渲染像素還會接收多個局部 SSIM 的梯度。

兩個簡化情形能進一步說明 D-SSIM 的尺度。若真實窗口只比渲染窗口多一個常數亮度偏移，即 $y_q=x_q+b$，其方差和協方差不變，因此

$$ 1-\operatorname{SSIM}(x,x+b) =\frac{b^2} {\mu_x^2+(\mu_x+b)^2+C_1}. $$

若只考察零均值窗口中的對比度縮放 $y=ax$，則

$$ 1-\operatorname{SSIM}(x,ax) =\frac{(a-1)^2\sigma_x^2} {(1+a^2)\sigma_x^2+C_2}. $$

亮度偏移 $b$ 和對比度偏差 $a-1$ 在匹配點附近都呈二次懲罰。L1 在非零小殘差處仍提供近似恆定幅值的梯度，D-SSIM 則更側重窗口統計失配；二者混合後形成了不同尺度的互補修正。

D-SSIM 在 3DGS 中的實際含義

3DGS 論文把結構損失記為 $\mathcal L_{\mathrm{D\text{-}SSIM}}$，並使用 [@kerbl2023gaussians]

$$ \mathcal L =(1-\lambda)\mathcal L_1 +\lambda\mathcal L_{\mathrm{D\text{-}SSIM}}, \qquad \lambda=0.2. $$

在 3DGS 的參考實現中，這一項直接計算為

$$ \mathcal L_{\mathrm{D\text{-}SSIM}} =1-\operatorname{SSIM}(\hat{\mathbf I},\mathbf I). $$

部分圖像處理文獻也把 D-SSIM 定義為 $(1-\operatorname{SSIM})/2$。兩種寫法相差常數倍率；3DGS 中的 $\lambda=0.2$ 與 $1-\operatorname{SSIM}$ 共同確定了結構項在混合目標中的尺度。D-SSIM 的定義發生縮放時，結構項相對 L1 的權重也會隨之改變。

混合目標的梯度組成

對任意高斯參數 $\theta_i$，混合目標的梯度可以展開為

$$ \frac{\partial\mathcal L}{\partial\theta_i} =(1-\lambda) \sum_p \frac{\partial\mathcal L_1}{\partial\hat{\mathbf I}(p)} \frac{\partial\hat{\mathbf I}(p)}{\partial\theta_i} +\lambda \sum_p \frac{\partial\mathcal L_{\mathrm{D\text{-}SSIM}}} {\partial\hat{\mathbf I}(p)} \frac{\partial\hat{\mathbf I}(p)}{\partial\theta_i}. $$

L1 部分直接從像素殘差給出修正方向。SSIM 部分會把一個像素與周圍窗口綁在一起。若 $s_p$ 是以 $p$ 為中心的局部 SSIM，$\mathcal W_p$ 是相應窗口，那麼渲染像素 $q$ 的結構梯度包含所有覆蓋它的窗口：

$$ \frac{\partial\mathcal L_{\mathrm{D\text{-}SSIM}}} {\partial\hat I_q} =-\frac{1}{|\Omega_s|} \sum_{p:,q\in\mathcal W_p} \frac{\partial s_p}{\partial\hat I_q}. $$

高斯對某個像素的貢獻由其屏幕 footprint 和透射率決定；SSIM 又把這個像素同鄰域統計相連。結構梯度經兩層局部支撐域傳回高斯，使邊緣、紋理和低對比度平面中的更新不再完全由單點顏色殘差決定。

Alpha 混合的梯度傳遞

對某個像素，把高斯按深度從前到後排序。令

$$ T_k=\prod_{j<k}(1-\alpha_j) $$

表示到達第 $k$ 個高斯前的累積透射率，$\mathbf B_{k+1}$ 表示從第 $k+1$ 個高斯開始、包含背景色的後景合成結果。像素顏色可在第 $k$ 層拆成

$$ \hat{\mathbf C} =\mathbf C_{<k} +T_k\left[ \alpha_k\mathbf c_k +(1-\alpha_k)\mathbf B_{k+1} \right]. $$

由此可以直接求出

$$ \frac{\partial\hat{\mathbf C}}{\partial\mathbf c_k} =T_k\alpha_k\mathbf I, \qquad \frac{\partial\hat{\mathbf C}}{\partial\alpha_k} =T_k(\mathbf c_k-\mathbf B_{k+1}). $$

第一式表明，顏色梯度由可見性 $T_k$ 和當前 Alpha 共同縮放。第二式則揭示了不透明度的可辨識條件：當高斯顏色與其後景接近時，改變 $\alpha_k$ 幾乎不改變像素顏色；當前景已令 $T_k$ 很小時，後方高斯同樣難以收到有效梯度。因此，某些高斯參數更新緩慢，未必是像素損失已經很低，也可能是當前視角中的合成雅可比接近退化。

忽略實現中的數值截斷，若屏幕空間高斯寫成

$$ G_k(\mathbf p) =\exp\left[ -\frac{1}{2}(\mathbf p-\boldsymbol\mu_k')^{\mathsf T} \mathbf Q_k (\mathbf p-\boldsymbol\mu_k') \right], \qquad \alpha_k=o_kG_k(\mathbf p), $$

其中 $\mathbf Q_k=(\boldsymbol\Sigma_k')^{-1}$，則投影中心的梯度為

$$ \frac{\partial\alpha_k}{\partial\boldsymbol\mu_k'} =o_kG_k(\mathbf p)\mathbf Q_k (\mathbf p-\boldsymbol\mu_k'). $$

再與 $\partial\mathcal L/\partial\alpha_k$ 相乘並對 footprint 內像素求和，就得到投影中心的更新方向。殘差分佈若在高斯兩側近似對稱，位置梯度會相互抵消；若殘差集中在某一側，中心便沿由 $\mathbf Q_k$ 調整過的方向移動。各向異性協方差由此同時影響覆蓋範圍和位置梯度的方向尺度。

$\lambda=0.2$ 也不能解讀成「結構梯度佔更新量的 20%」。兩項損失具有不同的數值尺度、空間相關性和梯度範數。某次迭代中結構項相對 L1 的實際影響更接近

$$ r= \frac{\lambda\left|\nabla_{\Theta} \mathcal L_{\mathrm{D\text{-}SSIM}}\right|2} {(1-\lambda)\left|\nabla{\Theta}\mathcal L_1\right|_2}. $$

即使 $\lambda$ 固定，$r$ 也會隨視角、紋理、當前殘差和高斯可見性變化。損失權重控制的是目標函數中的係數，實際參數更新還受到梯度形狀與 Adam 歷史動量影響。

這組對比還說明，單一損失曲線無法完整描述重建結果。訓練後期的標量損失可能只小幅下降，改善卻集中在小面積背景、細杆件或高頻紋理中；相反，平均損失下降也可能主要來自大面積平滑區域，而視覺上醒目的邊界仍有錯位。評價需要同時保留像素、結構和感知三個觀察尺度。

目標函數之外的優化結構

原論文的消融實驗保持重建損失不變，分別關閉初始化、密度控制、球諧外觀或梯度路徑。這組數據能夠區分「使用什麼損失」與「損失可以通過哪些參數和圖元傳播」 [@kerbl2023gaussians]。

在 $30\mathrm{k}$ 迭代時，移除 Split 比完整模型低 $2.15,\mathrm{dB}$，移除 Clone 只低 $0.14,\mathrm{dB}$，但圖中的細輻條已顯示出局部差異。單個場景的小面積錯誤在三場景平均 PSNR 中容易被稀釋，這也是圖像、逐場景數據和總平均需要同時保留的原因。

限制只有前 $10$ 個高斯接收梯度時，$30\mathrm{k}$ 平均 PSNR 下降 $6.86,\mathrm{dB}$，幾乎沒有從延長訓練中獲益。問題並不來自 L1 或 D-SSIM 定義，而在於遮擋後方的參數沒有完整收到誤差信號。損失數值、參數化與反向傳播實現因而共同決定實際優化問題。

四項常用重建指標

3DGS 論文使用 PSNR、SSIM 和 LPIPS 評價留出測試視角，實驗管線中也常加入 L1，便於直接觀察平均絕對殘差 [@kerbl2023gaussians; @chen2024survey]。

PSNR 與 MSE

對 $N$ 個像素通道，均方誤差為

$$ \operatorname{MSE} =\frac{1}{N}\sum_{j=1}^{N}(\hat I_j-I_j)^2. $$

峰值信噪比（Peak Signal-to-Noise Ratio，PSNR）定義為

$$ \operatorname{PSNR} =10\log_{10}\frac{\operatorname{MAX}^2}{\operatorname{MSE}}, $$

其中 $\operatorname{MAX}$ 是像素動態範圍上界。圖像歸一化到 $[0,1]$ 時，$\operatorname{MAX}=1$。固定動態範圍後，單幅圖像的 PSNR 是該圖 MSE 的單調變換，因而逐圖排序一致；測試集層面的排序還會受到後文所述聚合順序影響。

兩個方法的 PSNR 差值還能直接還原成 MSE 比值：

$$ \operatorname{PSNR}_A-\operatorname{PSNR}B =10\log{10}\frac{\operatorname{MSE}_B} {\operatorname{MSE}_A}. $$

因此，PSNR 提高約 $3.01\ \mathrm{dB}$，對應 MSE 約減半；提高 $1\ \mathrm{dB}$，對應 MSE 乘以約 $0.794$。dB 差值比孤立的 PSNR 數字更容易轉換成誤差幅度。

PSNR 對 MSE 的導數為

$$ \frac{\mathrm d\operatorname{PSNR}}{\mathrm d\operatorname{MSE}} =-\frac{10}{\ln 10}\frac{1}{\operatorname{MSE}}. $$

同樣大小的 MSE 絕對下降，在低誤差階段會換來更大的 dB 增量。因而 PSNR 曲線的縱向變化不能直接當作像素誤差的線性變化；比較訓練早期和後期時，回到 MSE 比值更容易判讀實際改善幅度。

L1 與 MSE 的界限

若圖像已歸一化到 $[0,1]$，則 $|e_j|\leq1$，有

$$ e_j^2\leq|e_j|. $$

對所有像素取平均，再利用均方根不小於算術平均，可得

$$ \operatorname{MSE} \leq\operatorname{L1} \leq\sqrt{\operatorname{MSE}}. $$

這組界限只約束數值範圍，不保證排序一致。少量大殘差會被平方項顯著放大，因而可能出現 L1 更低、PSNR 反而更差的情況。遇到反光、高亮、遮擋邊界或少量嚴重錯配時，這種指標分歧尤其常見。

微小位移下的圖像誤差

新視角中的幾何偏差常先表現為屏幕空間位移。假設渲染結果相對真實圖像存在一個很小的二維偏移 $\boldsymbol\delta$，一階泰勒展開給出

$$ \hat I(\mathbf p) =I(\mathbf p+\boldsymbol\delta) \approx I(\mathbf p) +\nabla I(\mathbf p)^{\mathsf T}\boldsymbol\delta. $$

相應的 L1 和 MSE 近似為

$$ \mathcal L_1 \approx\frac{1}{N}\sum_{\mathbf p} \left|\nabla I(\mathbf p)^{\mathsf T}\boldsymbol\delta\right|, $$

$$ \operatorname{MSE} \approx \boldsymbol\delta^{\mathsf T} \left[ \frac{1}{N}\sum_{\mathbf p} \nabla I(\mathbf p)\nabla I(\mathbf p)^{\mathsf T} \right] \boldsymbol\delta. $$

方括號內正是圖像梯度的二階矩陣。平滑區域的 $\nabla I$ 很小，幾何已發生位移也可能只產生有限像素誤差；單一方向的邊緣主要約束其法向位移；角點和豐富紋理在兩個方向上都能提供較強約束。這解釋了弱紋理平面容易獲得不錯的 RGB 指標，卻仍可能保留深度或位置偏差，也說明了細杆件與高對比邊界為何對亞像素錯位格外敏感。

LPIPS 的特徵距離

學習感知圖像塊相似度（Learned Perceptual Image Patch Similarity，LPIPS）把兩幅圖像送入固定的深度網絡，在多層特徵上計算加權距離 [@zhang2018lpips]。省略實現細節後，可寫成

$$ \operatorname{LPIPS}(x,y) =\sum_l\frac{1}{H_lW_l} \sum_{h,w} \left| \mathbf w_l\odot \left( \hat{\phi}l(x){h,w} -\hat{\phi}l(y){h,w} \right) \right|_2^2, $$

其中 $\hat{\phi}_l$ 表示第 $l$ 層歸一化特徵，$\mathbf w_l$ 是學習得到的通道權重。LPIPS 允許像素坐標上存在一定差異，只要深層特徵仍然接近；這使它能補充觀察紋理與局部外觀，也令結果依賴骨幹網絡、權重版本和輸入歸一化方式。

通道歸一化賦予了這個距離一個直觀的幾何解釋。若暫時忽略學習權重，並令兩個特徵向量 $\hat{\mathbf f}$、$\hat{\mathbf g}$ 的二範數均為 $1$，則

$$ \left|\hat{\mathbf f}-\hat{\mathbf g}\right|_2^2 =2-2\hat{\mathbf f}^{\mathsf T}\hat{\mathbf g} =2\left(1-\cos\vartheta\right). $$

未加權的局部特徵距離等價於比較兩個特徵方向的餘弦相似度。加入 $\mathbf w_l$ 後，距離變為

$$ \left(\hat{\mathbf f}-\hat{\mathbf g}\right)^{\mathsf T} \operatorname{diag}(\mathbf w_l^2) \left(\hat{\mathbf f}-\hat{\mathbf g}\right), $$

即在特徵空間中使用學習得到的對角度量。權重較大的通道對最終分數影響更強，而歸一化削弱了整體激活幅值的作用。LPIPS 因此測量的是經骨幹網絡與感知數據共同塑形的特徵差異，並非與網絡無關的通用視覺距離。

LPIPS 較低不等於幾何一定更準確。模糊、微小位移、重複紋理或網絡訓練域之外的材質，都可能令感知特徵與幾何誤差之間出現偏差。涉及三維幾何的研究仍需另行報告深度、法線、點雲或網格指標。

評價指標的聚合方式

假設測試集包含 $K$ 幅圖像，每幅圖像的 MSE 為 $m_k$。常見的「逐圖計算 PSNR 後取平均」為

$$ \overline{\operatorname{PSNR}} =\frac{1}{K}\sum_{k=1}^{K} 10\log_{10}\frac{\operatorname{MAX}^2}{m_k}, $$

而「先合併全部像素的 MSE 再換算」近似為

$$ \operatorname{PSNR}{\mathrm{global}} =10\log{10} \frac{\operatorname{MAX}^2} {\frac{1}{K}\sum_{k=1}^{K}m_k}. $$

把 $m_k$ 的算術平均和幾何平均分別記為

$$ M_{\mathrm A}=\frac{1}{K}\sum_{k=1}^{K}m_k, \qquad M_{\mathrm G}=\left(\prod_{k=1}^{K}m_k\right)^{1/K}, $$

則兩種 PSNR 可分別寫成

$$ \overline{\operatorname{PSNR}} =10\log_{10}\frac{\operatorname{MAX}^2}{M_{\mathrm G}}, \qquad \operatorname{PSNR}{\mathrm{global}} =10\log{10}\frac{\operatorname{MAX}^2}{M_{\mathrm A}}. $$

二者的差值恰好為

$$ \overline{\operatorname{PSNR}} -\operatorname{PSNR}{\mathrm{global}} =10\log{10}\frac{M_{\mathrm A}}{M_{\mathrm G}} \geq0. $$

最後一步來自算術平均不小於幾何平均。這個差值不只說明兩種口徑的方向，也刻畫了逐圖 MSE 的離散程度：各視角誤差完全相同時差值為零，難易視角分化越明顯，$M_{\mathrm A}/M_{\mathrm G}$ 通常越大。

因此，兩者滿足

$$ \overline{\operatorname{PSNR}} \geq\operatorname{PSNR}_{\mathrm{global}}. $$

若各圖像分辨率不同，先逐圖平均還會讓小圖和大圖具有相同權重；全像素聚合則由大圖主導。SSIM、L1 和 LPIPS 也存在逐圖平均、逐場景平均、按像素加權等口徑差異。報告結果時，聚合方法應當與數據劃分、圖像縮放和掩碼規則一起固定。

訓練損失與評價指標

同一方法內部，訓練損失曲線可以用來觀察是否收斂、是否出現數值異常，以及某次密度控制後優化狀態如何恢復。跨方法比較時，只有損失定義、權重、掩碼、歸一化和採樣方式完全一致，總損失數值才有直接比較意義。

若方法 $A$ 的目標為

$$ \mathcal L_A =0.8\mathcal L_1+0.2\mathcal L_{\mathrm{D\text{-}SSIM}}, $$

方法 $B$ 在此基礎上加入正則項

$$ \mathcal L_B =\mathcal L_A+\beta\mathcal L_{\mathrm{reg}}, $$

那麼 $\mathcal L_B>\mathcal L_A$ 不能直接推出方法 $B$ 的 RGB 重建更差。額外項改變了標量基準，也可能犧牲少量訓練視角的像素擬合，換取更穩定的幾何或更好的留出視角。這類比較應拆開查看共同損失分量，再使用同一測試劃分、同一渲染設置和同一指標實現評價最終結果。

SSIM 同時出現在訓練和評價中也需要分開記錄。訓練時它是帶梯度的局部結構約束，只在當前採樣視角上參與更新；評價時模型保持固定，SSIM 在留出視角上統計泛化結果。即使兩處調用同一段函數，觀察對象仍不相同。

重建評價的實驗設置

指標表之外，至少需要固定以下條件：

• 測試視角：使用相同的訓練／測試劃分，避免把參與優化的視角混入評價。
• 相機與分辨率：保持相同內參、裁剪、縮放和渲染尺寸；亞像素偏移也會同時影響四項指標。
• 顏色空間：明確在線性 RGB 還是 sRGB 中計算，並統一像素動態範圍。
• 背景與 Alpha：透明區域應使用相同背景合成；前景掩碼的使用範圍也要一致。
• 指標實現：記錄 SSIM 的窗口與邊界處理、LPIPS 的骨幹和權重版本，以及各指標的聚合順序。
• 可視化：保存同一視角的真實圖像、渲染圖像和誤差圖，優先檢查高亮、遮擋邊界、細杆件、重複紋理與弱紋理平面。
• 效率條件：涉及訓練時間、幀率或顯存時，固定硬件、圖像尺寸、高斯數量統計時刻和渲染設置。

對每個測試視角保留逐圖指標，通常比只保存場景均值更有價值。均值回答整體趨勢，逐圖分布可以定位失敗視角；中位數、分位數或最差若干視角則能補充平均值容易掩蓋的長尾問題。

結語

3DGS 的混合訓練目標把兩種誤差信號送入同一條可微渲染鏈：L1 提供穩定、直接的逐像素顏色修正，D-SSIM 通過局部統計把相鄰像素聯繫起來。梯度到達高斯參數之前，還要經過透射率、前後景顏色差和屏幕空間高斯的雅可比；損失權重相同，不代表每個高斯能接收到同等強度的更新。

重建評價則從不同尺度觀察留出視角：L1 和 PSNR 描述像素殘差，SSIM 觀察局部結構，LPIPS 補充深度特徵距離。微小位移的展開表明，RGB 指標的敏感度本身受圖像梯度方向支配；PSNR 的算術平均—幾何平均差值又會放大視角難度分化。任何單項指標都有盲區，總損失也只有在定義一致時才能橫向比較。將公式、實現配置、聚合口徑和同視角圖像放在一起，指標才真正具有可解釋性。

@article{kerbl2023gaussians,
  author = {Kerbl, Bernhard and Kopanas, Georgios and Leimkühler, Thomas and Drettakis, George},
  title = {3D Gaussian Splatting for Real-Time Radiance Field Rendering},
  journal = {ACM Transactions on Graphics},
  volume = {42},
  number = {4},
  pages = {139:1--139:14},
  year = {2023}
}

@article{chen2024survey,
  author = {Chen, Guikun and Wang, Wenguan},
  title = {A survey on 3d gaussian splatting},
  journal = {ACM Computing Surveys},
  year = {2024}
}

@article{wang2004ssim,
  author = {Wang, Zhou and Bovik, Alan C. and Sheikh, Hamid R. and Simoncelli, Eero P.},
  title = {Image Quality Assessment: From Error Visibility to Structural Similarity},
  journal = {IEEE Transactions on Image Processing},
  volume = {13},
  number = {4},
  pages = {600--612},
  year = {2004}
}

@inproceedings{zhang2018lpips,
  author = {Zhang, Richard and Isola, Phillip and Efros, Alexei A. and Shechtman, Eli and Wang, Oliver},
  title = {The Unreasonable Effectiveness of Deep Features as a Perceptual Metric},
  booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages = {586--595},
  year = {2018}
}

Since I haven’t assembled my own MiSTer setup yet, all information here is currently sourced from the internet and will be updated once I have hands-on experience.

:::

Retro gaming is often sold as a content problem: find the ROM, map the buttons, press Start, and the past reappears on demand.

MiSTer refuses that framing. It treats retro gaming as a behavior problem—a question of clocks, buses, contention, odd refresh rates, audio timing, and the small edge cases that were never written down because the original machines were the documentation. When those machines age, fail, and disappear into parts bins, the thing we lose is not only plastics and PCBs. We lose a way of working.

MiSTer’s most honest description is not “a device,” but a platform: a shared hardware baseline plus an open toolchain and a living library of FPGA “cores” that can reconstruct many classic computers, consoles, and arcade boards on modern programmable logic. It is messy in the way real preservation work is messy. It asks you to care about the chain rather than the screenshot. And if you accept that bargain, it can turn “I still want to play this” into something closer to “this remains operable.”

A platform, not a box

A typical MiSTer build starts with the Terasic DE10-Nano: an FPGA development board that pairs an Intel Cyclone V FPGA with an ARM-based SoC running a lightweight Linux environment. The Linux side handles the everyday conveniences—menus, storage, networking, scripts, controllers—while the FPGA side loads a core that implements the target system’s logic.

That split is the key to why MiSTer feels different from both emulators and consumer FPGA products. It’s not trying to hide its layers. The “main” MiSTer software is not a single monolithic game launcher; it is the host runtime that knows how to configure the FPGA, expose a common interface, and coordinate video/audio/input across wildly different machines. The cores are the exhibits, and the platform is the cabinet that keeps them accessible.

This is also why MiSTer doesn’t have one stable “product identity.” Your MiSTer is partly defined by the hardware you stack onto the DE10-Nano, partly by the updater workflow you trust, partly by the display chain you build around it, and partly by the cores you actually use. In practice it becomes less like buying a console and more like adopting an ecosystem.

“Core” is not a ROM slot: it is a behavioral claim

MiSTer cores are FPGA implementations of classic hardware, written in hardware description languages and built within a shared framework. When a core is loaded, the FPGA is configured to behave like the target system’s components and their relationships—CPU timing, video pipeline, audio generation, memory mapping, bus arbitration, and the subtle “this register does something odd if you read it twice” realities that define many real machines.

That does not mean every core is perfect, and MiSTer is at its best when it admits this openly. Accuracy is not a binary property; it is a gradient, and it can vary by subsystem. A core can be “good enough to finish the game” while still being wrong in ways that matter for speedrunners, for demoscene software, or for the small set of titles that lean on undocumented behavior. MiSTer’s value is that it makes these questions discussable in engineering terms: a bug is something you can reproduce, isolate, patch, and version—not a vague feeling that your laptop is “doing something.”

This also reframes what “compatibility” means. In emulator culture, compatibility is often judged by whether a title boots and plays. In MiSTer culture, compatibility tends to drift toward whether a system’s behavioral envelope is represented: timing edge cases, video modes, audio quirks, and the interactions between them. It’s not that MiSTer users are always more demanding; it’s that the platform invites a different standard of argument.

Why the SDRAM module exists (and why it keeps showing up in every build list)

One of MiSTer’s most confusing practical details is that the DE10-Nano already has DDR3 memory, and yet many builds add an external SDRAM module anyway.

The short explanation is latency and determinism. A significant number of cores rely on low-latency external SDRAM because the onboard DDR3 is not optimized for the kinds of timing-sensitive access patterns some classic systems expect. MiSTer’s own hardware documentation frames the external SDRAM module as a solution for cores that require larger, fast memory, noting that the DE10-Nano’s DDR3 has high latency for this use case.

This is a perfect example of MiSTer’s philosophy leaking into the bill of materials. If you only care whether a game draws the right pixels, you can hand-wave memory timing. If you care whether a system behaves like itself across its weird corners, memory behavior stops being “just RAM.” It becomes part of the system’s identity.

The common outcome is practical: many people treat SDRAM not as an optional upgrade but as the first “you probably want this” addition, because it expands the set of cores you can run comfortably and reduces the number of platform-specific caveats you’ll hit later.

The display chain is half the experience

MiSTer’s reputation is often summarized as “low latency” or “hardware accurate,” but what many users are actually reacting to is something more mundane and more important: MiSTer encourages you to treat the display path as part of the machine.

Classic consoles and arcade boards did not target HDMI. They targeted CRT behaviors: 240p, 480i, odd refresh rates, non-square pixels, and signal characteristics that modern flat panels often “correct” in ways that produce added lag, unstable motion cadence, or scaling artifacts. Software emulation can absolutely look fantastic, but it is typically mediated by a modern OS compositor and a modern display pipeline that was never designed to respect those old assumptions.

MiSTer offers multiple strategies, and the best one depends on what you are trying to preserve.

If you are primarily on modern displays, MiSTer’s HDMI output gives you a controllable, consistent source where you can tune scaling and sync behavior per core, and you can keep the system’s timing intent more intact than many generic “retro modes” on televisions.

If you are CRT-minded—or simply want to keep analog output as a first-class option—MiSTer supports analog video in two culturally important ways.

One path is the classic I/O board approach, where an add-on board provides analog outputs designed for CRT and other legacy equipment.

The other is MiSTer’s “Direct Video” mode, which routes an analog-compatible signal through the HDMI port using a suitable DAC/adapter. The MiSTer documentation describes Direct Video as an option that avoids requiring an I/O board while still delivering minimal latency comparable to the I/O board’s VGA output, and in some cases offering improved color depth depending on the core. In the MiSTer world, that’s not a footnote; it’s an ideological statement: analog is not just a legacy connector, it is a preserved grammar of display.

The consequence is that MiSTer tends to produce fewer “it looks right but feels wrong” moments, because you are encouraged—almost forced—to think about the entire path from button press to phosphor glow.

Latency: where MiSTer helps, and where it can’t

MiSTer cannot repeal physics. It can, however, reduce the number of layers that lie to you.

A modern gaming PC running an emulator can be extremely fast and extremely accurate, yet still accumulate latency through a chain of small decisions: OS scheduling, input polling, buffering, frame pacing, GPU queues, scaling, display processing. Any one of those may be negligible, but the sum can be felt—especially in genres that were built around immediate feedback.

MiSTer’s architecture tends to simplify the critical loop. Input is read by the host environment and delivered into an FPGA core that generates frames and audio with deterministic timing goals. That does not guarantee “zero lag,” but it does reduce the amount of “who knows what happened between my controller and the game.”

The bigger enemy often becomes the display. Many modern televisions add processing, frame buffers, motion interpolation logic, and scaling pipelines unless you explicitly disable them. MiSTer can deliver a clean signal with consistent cadence; it cannot force your panel to respect it.

This is why MiSTer culture frequently sounds like CRT culture even when the user is not a CRT purist. The point is not nostalgia for glass. The point is respect for timing.

Maintenance is not a downside; it is the price of keeping things operable

MiSTer is sometimes marketed as “pay once and get accuracy.” That is not how it behaves.

MiSTer behaves like a maintained system because it is one. Cores change, frameworks evolve, and community tooling moves. MiSTer’s official Distribution repository exists precisely because the platform is more than a single binary; it is a coherent set of files that make an SD card “a MiSTer.” The MiSTer Downloader tool exists because keeping up-to-date is not a side feature; it is essential infrastructure. The Downloader README describes its scope bluntly: it installs and updates cores and extra files, and it also updates the menu core, MiSTer firmware, and even the Linux system, pulling from the Distribution repository.

Then there is the reality that most MiSTer users eventually adopt a higher-level script workflow. The widely used “Update All” script explicitly positions itself as an all-in-one updater that runs MiSTer Downloader under the hood and expands it with additional databases and options. Even its own README includes a warning that scripts run with root access and should be treated with caution—a reminder that MiSTer is closer to a small computer you administer than a console you consume.

This maintenance aspect is not an accident. It is the inevitable cost of a platform that wants to outlive hardware scarcity. If you want a sealed appliance, you will be happier elsewhere. If you want a system that stays current because a community is still actively making it better, MiSTer’s upkeep is part of the deal.

Can MiSTer run original cartridges?

This is the point where many people’s mental model diverges, especially if they are coming from a world where “FPGA hardware” is marketed as “use your original collection.”

In theory, running original cartridges is “just I/O”: a physical ROM device on a bus with timing requirements. In practice, MiSTer is not designed to be a universal cartridge host, and the project’s own documentation is explicit about the direction here. The MiSTer FAQ states that MiSTer will never officially use physical cartridges, describing cartridge support as physically impractical given the number of GPIO pins available from the FPGA, and framing the project goal as replacing the need for original hardware rather than becoming an accessory dock for it.

That does not mean your original cartridges are irrelevant. It simply means MiSTer treats them as sources, not as runtime media. The common preservation workflow is to dump a cartridge (using appropriate external hardware) into a ROM image and then run that image from MiSTer’s storage like any other title. Conceptually, the cartridge remains part of the provenance of the game; operationally, MiSTer standardizes around files to keep the platform manageable and broadly compatible.

This detail also clarifies an important distinction. MiSTer can be very friendly to “original-feeling” peripherals—controllers, light guns (with the right display constraints), and so on—because those map to narrower, more controllable interfaces. Physical game media is a much harder universal target, and MiSTer chooses not to make that trade.

MiSTer and openFPGA: separate ecosystems, shared DNA

MiSTer is frequently mentioned alongside Analogue’s openFPGA, and at a distance the resemblance is obvious. Both involve FPGA cores. Both attract developers who care about authenticity. Both encourage a “hardware first” mental model.

But the relationship is not direct compatibility. It is closer to a family resemblance.

Analogue’s developer documentation defines a Pocket “core” as an FPGA bitstream packaged with JSON definition files, with loading and operation managed by the Analogue Platform Framework (APF). The APF provides abstractions for buses and I/O, and Analogue’s documentation describes how integration code simplifies access to BRIDGE and PAD buses while providing examples for VIDEO and AUDIO.

MiSTer cores, by contrast, live in a different host environment with different expectations, different file structures, and a different runtime framework. Even when two cores implement the same target system, they are not drop-in equivalents. A port typically requires rebuilding the platform glue: how assets are loaded, how input is surfaced, how video modes are negotiated, how saves are handled, and how the core’s assumptions map onto the host’s bus abstractions.

This is why you can truthfully say “many openFPGA cores are ports of MiSTer work” while also truthfully saying “MiSTer cores are not directly runnable on Pocket.” The overlap tends to be in upstream HDL designs and developer communities, not in binary compatibility.

Cross-platform developers make that boundary feel thinner than it is. When one developer maintains similar implementations across both ecosystems, the user experience can converge—even though the frameworks remain distinct.

In short: MiSTer and openFPGA are cousins, not clones. MiSTer optimizes for openness and modularity as a community platform. openFPGA optimizes for a developer framework inside a product ecosystem. Both can serve preservation; they simply do it through different kinds of sustainability.

What MiSTer is really preserving

MiSTer is not preserving “games” in the narrow sense. It is preserving a way of making games run—a set of electrical and temporal assumptions that were once ordinary and are now exotic.

That matters because original hardware is becoming less maintainable, not more. The longer we wait, the more “authentic play” risks becoming an artifact of who can source parts, who has repair knowledge, and who can tolerate failure. MiSTer does not replace original hardware’s material history, but it can prevent operability from becoming a luxury.

There is also a forward-looking implication that makes MiSTer feel larger than “retro.” A well-maintained core is, in effect, a publishable artifact: a versioned, reviewable, reproducible behavioral model of a classic machine. That is a preservation strategy with a future. It treats hardware not as a finite stash of aging boards, but as something that can be documented in logic and kept alive as long as we still know how to program logic.

MiSTer is not the only path to that future, but it is one of the clearest statements that the past is not just something we watch. It is something we can keep operable—if we are willing to treat it seriously.

References

MiSTer (official / primary)

• MiSTer FPGA Documentation (MkDocs): https://mister-devel.github.io/MkDocs_MiSTer/

• MiSTer-devel organization (core and platform repos): https://github.com/MiSTer-devel

• MiSTer Hardware official repo: https://github.com/MiSTer-devel/Hardware_MiSTer

MiSTer (community tooling)

• Update All (theypsilon): https://github.com/theypsilon/Update_All_MiSTer

openFPGA (Analogue primary docs)

• Analogue Developer Docs : https://www.analogue.co/developer/docs/overview

• openFPGA Library : https://openfpga-library.github.io/analogue-pocket/

• Jotego jtcores (multi-platform intent, incl. MiSTer / Analogue Pocket): https://github.com/jotego/jtcores

• Analogue “Getting Started with openFPGA”: https://www.analogue.co/support/resource/getting-started-with-openfpga

「乾杯是日語，干杯是和製漢語，是日本先形成了『乾杯』作為祝酒詞，之後再傳回中國。」

當這一說法遭到質疑時，往往又會補充一個看似合理的疑問：中日兩邊在近代幾乎同時把「干杯／乾杯」用作祝酒口號，這是否只是巧合？

下文試圖從語言學的角度說明：這一現象與其說是 lexical borrowing（詞彙借入）或 accidental similarity（偶然相似），不如理解為一種在既定條件下高度可預期的 parallel development（並行演化）。換言之，所謂「巧合」本身，可能正是語言在相似約束下自然收斂的結果。

書寫系統與詞彙來源的區分

討論首先需要釐清一個在日常語言討論中經常被混用的區分：orthographic inheritance（書寫系統的繼承）並不等同於 lexical origin（詞彙起源）。

日語中的「乾杯（かんぱい）」屬於典型的 Sino-Japanese lexeme，其特徵包括：以漢字書寫、僅有 On-yomi（音讀）而無 Kun-yomi（訓讀），其讀音層主要可追溯至中古漢語。從 linguistics 的角度看，這樣的事實只能說明該詞的 phonological layer 來自漢語輸入，卻不足以支持詞彙是在日語中被創造，或其 pragmatic function 曾由日語反向影響漢語。

若僅依賴 On-yomi 來推斷詞源方向，則「革命」「文化」「哲學」「經濟」等大量音讀詞亦將面臨相同的解釋困境。這顯示，僅憑書寫與讀音形式，很難支撐「日本原創—再傳回」這一路徑。

構詞法與語義透明性

從構詞法與語義學的角度看，「干杯」本身是一個語義高度透明的構式：干／乾作為 causative 或 resultative 成分，表示「使之為空」，杯則指飲酒器具，其組合語義可直觀理解為 make the cup empty，使杯為空。

這種透明性意味著，該表達並不需要假設外來 borrowing 才能成立。事實上，在近代以前的漢語中，功能與結構相近的表達早已存在，如「飲盡此杯」「干一杯」「須得飲乾」等，尤其在明清白話文獻中，此類 exhortative drinking expressions 的使用頻率相當高。

從語言接觸研究的一般觀點來看，當一種語言已具備充分的 internal resources 來表達某一語義與語用功能時，lexical borrowing 通常並非最自然的解釋路徑。

歷史語用學層面的觀察

不少相關討論隱含了一個前提假設：古代可能早已有一個固定的祝酒口號，只是後來被不同語言採用。然而從 historical pragmatics 的角度來看，中日古代的飲酒語言都呈現出相似特徵：ritualized action 在交際中的地位高於 verbal formula，與飲酒相關的 speech 多以完整句或祝辭形式出現，而非短促、可齊聲呼喊的 exclamatory formula。

早期文獻中出現的「乾杯」，更接近 action-oriented expression，而非 discourse marker。這一點使得「某一方較早完成口號化」的假設缺乏必要的歷史支撐。

接觸誘發的語用再功能化

若將注意力從「詞從何來」轉向「功能如何形成」，問題便轉化為：為何在近代語境中，中日語言都逐漸需要一個 toast-initiating 的言語行為？

在語言學中，這一過程可置於 ++contact-induced pragmatic refunctionalization++ 的框架下理解。簡而言之，既有構式作為 pre-existing construction 長期存在；19 世紀以後，interactional ecology 因西式宴會文化、外交活動以及現代組織形態而發生變化；原有表達被重新分配為 ritualized speech act；並最終經由 pragmatic conventionalization 固定為 discourse-level exclamation。

在這一分析框架中，並不需要引入 lexical borrowing 作為前提。

為何不宜簡化為「偶然巧合」

在語言學中，accidental similarity 應是最後的解釋選項，其成立通常要求：

• languages are unrelated（語言彼此無關）

• structures are non-transparent（結構不透明）

• 不存在 shared sociocultural trigger（共享的社會文化觸發因素）

而「干杯／乾杯」的情況恰恰相反：中日共享漢字書寫系統與漢語構詞體系，變化前具有相同的 semantic baseline，近代又同時接觸西方祝酒慣例，且最終結果在用法、語氣與限制上並非完全一致，呈現的是 functional convergence 而非 copying。在此條件組合下，parallel development是最符合 ++簡約性原則++ 的解釋。

反向檢驗與語言經濟

從語言經濟的角度看，語言變化往往遵循某種 least-effort principle。當構式已存在、語義已對齊，而新的交際需求突然出現時，「干杯／乾杯」被選中，與其說是巧合，不如說是一種預期之內的結果。反而是「若不用它，為何另造他詞」，才需要進一步的理論說明。

結論

將「干杯／乾杯」理解為「日語傳回中文」，在語言學分析中更接近一種 category error，即把 pragmatic conventionalization 誤認為 lexical borrowing。

一種更為穩健的理解是：在高度受約束的語言與社會條件下，中日語言分別完成了 parallel pragmatic refunctionalization。這一結果並非巧合，也不必訴諸單向傳播，而是語言在現代語境中自然演化的一種表現。

給定同一個靜態場景的多張照片，以及每張照片對應的相機內參和位姿，任務是在訓練數據之外生成新的觀察視角。這個問題稱為新視角合成（Novel View Synthesis，NVS）。完成這項任務需要先建立可供渲染的場景表示，再依照目標相機的位置與方向生成圖像。

NeRF 使用神經網絡表示連續的輻射場。渲染目標視角時，每個像素向場景發出一條光線，沿光線採樣多個空間位置，查詢各位置的顏色與密度，最後以體積渲染累積成像。這條路徑具有很強的表示能力，高分辨率圖像卻會帶來大量空間採樣與網絡推理。

3D Gaussian Splatting（以下簡稱 3DGS）沿用多視角圖像監督，場景表示改為一組可學習的三維高斯。渲染時，高斯由三維空間投影到目標相機的圖像平面，再經光柵化與 Alpha 混合合成像素。NeRF 的計算順序是「像素—光線採樣—場景查詢」，3DGS 則是「三維圖元—圖像投影—像素合成」；後者更適合 GPU 並行處理，也由此獲得實時渲染能力 [@kerbl2023gaussians; @chen2024survey]。

原論文的總覽圖把這條路徑連同訓練時的梯度反傳放在同一張圖中：SfM 點雲提供初始幾何，高斯經投影與光柵化形成圖像，圖像誤差再經反向傳播送回投影、密度控制與高斯參數。

3DGS 將場景表示的基本單元直接納入優化：每個高斯的參數會連續更新，高斯數量也會隨訓練增減。渲染器在生成圖像的同時，還需要為局部基底的移動、縮小和增生提供梯度信號。後文的公式推導將沿著這條梯度鏈展開。

三維高斯的場景表示

3DGS 使用 SfM（Structure from Motion）估計相機位姿，並利用 SfM 附帶生成的稀疏點雲初始化高斯中心。對第 $i$ 個高斯，可以把其狀態概括為

$$ \mathcal{G}_i = \left(\mu_i,,\mathbf{s}_i,,\mathbf{q}_i,,\alpha_i,,\mathbf{h}_i\right). $$

各部分的含義如下：

若使用三階球諧，也就是 $L=3$，單個高斯以浮點數實際存儲的參數量為

$$ 3_{\mu}+3_{s}+4_{q}+1_{\alpha} +3(L+1)^2_{\mathrm{SH}} =59. $$

四元數雖然存儲四個量，歸一化後只有三個旋轉自由度。以 FP32 計算，59 個量佔 236 bytes；一百萬個高斯僅模型參數就約為 225 MiB，尚未計入 Adam 的一、二階動量與光柵化器的臨時數據。這個簡單的估算已經說明，3DGS 的實時渲染速度不等於表示本身輕量，壓縮與結構化存儲會成為後續研究的重要分支。

以中心 $\mu_i$ 和協方差 $\Sigma_i$ 定義的三維高斯核為

$$ G_i(\mathbf{x})= \exp!\left[-\frac{1}{2} (\mathbf{x}-\mu_i)^\mathsf{T} \Sigma_i^{-1} (\mathbf{x}-\mu_i) \right]. $$

式中省略了機率密度函數的歸一化係數。3DGS 將高斯核用作具有空間範圍的軟圖元：核函數描述影響力從中心向外衰減的方式，可學習不透明度 $\alpha_i$ 控制其幅值。

協方差的尺度—旋轉參數化

梯度更新若直接作用於任意 $3\times3$ 矩陣，可能破壞協方差所需的半正定性。3DGS 分別學習尺度與旋轉，再由兩者構造 $\Sigma_i$：

$$ \mathbf{S}_i=\operatorname{diag}(\mathbf{s}_i), \qquad \Sigma_i =\mathbf{R}(\mathbf{q}_i)\mathbf{S}_i\mathbf{S}_i^\mathsf{T} \mathbf{R}(\mathbf{q}_i)^\mathsf{T}. $$

這種寫法便於實現，並從構造上保證了半正定性。對任意向量 $\mathbf{z}$，有

$$ \begin{aligned} \mathbf{z}^\mathsf{T}\Sigma_i\mathbf{z} &=\mathbf{z}^\mathsf{T}\mathbf{R}\mathbf{S}\mathbf{S}^\mathsf{T}\mathbf{R}^\mathsf{T}\mathbf{z}\ &=\left|\mathbf{S}^\mathsf{T}\mathbf{R}^\mathsf{T}\mathbf{z}\right|_2^2\ &\ge 0. \end{aligned} $$

只要尺度保持為正、旋轉四元數經過歸一化，這個矩陣就始終具有合理的橢球幾何意義。實現中令無約束變量 $\boldsymbol{\rho}_i\in\mathbb{R}^3$ 經指數函數映射為尺度，並以 Sigmoid 映射不透明度：

$$ \mathbf{s}_i=\exp(\boldsymbol{\rho}_i), \qquad \alpha_i=\operatorname{sigmoid}(o_i). $$

把 $\mathbf{R}i$ 的第 $k$ 個列向量記作 $\mathbf{r}{i,k}$，協方差還可以展開成

$$ \Sigma_i =\mathbf{R}i\operatorname{diag} !\left(e^{2\rho{i,1}},e^{2\rho_{i,2}},e^{2\rho_{i,3}}\right) \mathbf{R}i^\mathsf{T} =\sum{k=1}^{3}e^{2\rho_{i,k}} \mathbf{r}{i,k}\mathbf{r}{i,k}^{\mathsf T}. $$

所以它的三個特徵值正好是 $e^{2\rho_{i,k}}$，有限尺度下嚴格大於零；對數尺度的導數為

$$ \frac{\partial\Sigma_i}{\partial\rho_{i,k}} =2e^{2\rho_{i,k}} \mathbf{r}{i,k}\mathbf{r}{i,k}^{\mathsf T}. $$

這裡有一個很實用的含義：$\rho$ 的加法更新對應 $\mathbf{s}$ 的乘法縮放。不同大小的高斯可以共用相近量級的優化步長，梯度也始終沿著合法橢球的主軸調整。協方差參數化由此同時提供幾何約束與數值尺度的整理。

各向異性提高了表示的緊湊性。球形高斯只有一個尺度，覆蓋牆面時需要堆疊大量小球；可旋轉、可拉伸的橢球能貼近牆面、葉片、欄杆或物體邊界，用較少圖元描述具有方向性的局部結構。

原論文將收斂後的高斯縮小至原尺度的 60%，藉此顯露平時被 splatting 覆蓋的橢球輪廓。通風柵的細長葉片由方向一致的扁平高斯排列而成，直觀呈現了協方差對局部形狀的刻畫。

球諧係數與視角相關顏色

同一個位置從不同方向觀看時，顏色可能因高光等效應而改變。3DGS 為每個高斯保存球諧（Spherical Harmonics，SH）係數，將觀察方向 $\mathbf{d}$ 上的顏色寫成

$$ \mathbf{c}i(\mathbf{d})= \sum{\ell=0}^{L}\sum_{m=-\ell}^{\ell} \mathbf{h}{i,\ell m}Y{\ell m}(\mathbf{d}). $$

原文所稱的四個 SH bands 對應 $\ell=0,1,2,3$，因此每個顏色通道共有

$$ \sum_{\ell=0}^{3}(2\ell+1)=(3+1)^2=16 $$

個基底係數。訓練從零階的基礎顏色開始，再逐步加入更高階頻帶。這能降低角度觀測不足時，高階方向成分過早吸收誤差所造成的不穩定。

SH 描述的是外觀隨方向的變化。光源、材質與反射仍未被顯式建模，因此這些係數可以擬合一部分視角相關現象，卻不會自動給出物理正確的反照率或光照分解。

三維高斯的二維投影

渲染時，世界空間中的三維橢球會被投影成圖像平面上的二維橢圓。中心點遵循普通的相機投影，橢圓形狀則由協方差在透視投影下的變換決定。

設高斯中心經世界到相機變換後為

$$ \mathbf{x}c= \begin{bmatrix}x_c&y_c&z_c\end{bmatrix}^\mathsf{T} =\mathbf{R}{cw}\mu_i+\mathbf{t}_{cw}. $$

針孔相機的投影函數為

$$ \pi(\mathbf{x}_c)= \begin{bmatrix} f_x x_c/z_c+c_x\ f_y y_c/z_c+c_y \end{bmatrix}, $$

所以二維中心是 $\mu'_i=\pi(\mathbf{x}_c)$。透視除法屬於非仿射變換，協方差無法直接套用線性變換規則。3DGS 在高斯中心附近對 $\pi$ 做一階泰勒展開：

$$ \pi(\mathbf{x}_c+\Delta\mathbf{x}) \approx \pi(\mathbf{x}_c)+\mathbf{J}_i\Delta\mathbf{x}, $$

其中投影雅可比矩陣為

$$ \mathbf{J}_i= \begin{bmatrix} f_x/z_c & 0 & -f_xx_c/z_c^2\ 0 & f_y/z_c & -f_yy_c/z_c^2 \end{bmatrix}. $$

這個雅可比可以直接從透視除法的微分得到。以水平座標 $u=f_xx_c/z_c+c_x$ 為例，

$$ \mathrm{d}u =\frac{f_x}{z_c},\mathrm{d}x_c -\frac{f_xx_c}{z_c^2},\mathrm{d}z_c, $$

垂直座標同理。它也揭示了深度的作用：相同的三維位移在近處產生更大的像素位移，沿深度方向的擾動還會被 $x_c/z_c^2$ 或 $y_c/z_c^2$ 放大。

把高斯中心附近的世界空間偏移記為隨機向量 $\Delta\mathbf{x}_w$，滿足

$$ \mathbb{E}[\Delta\mathbf{x}_w]=\mathbf{0}, \qquad \mathbb{E}[\Delta\mathbf{x}_w\Delta\mathbf{x}_w^\mathsf{T}]=\Sigma_i. $$

相機空間偏移為 $\Delta\mathbf{x}c=\mathbf{R}{cw}\Delta\mathbf{x}_w$，一階投影偏移為 $\Delta\mathbf{p}\approx\mathbf{J}_i\Delta\mathbf{x}_c$。因此二維協方差可以逐步推出：

$$ \begin{aligned} \Sigma_i^{\prime} &=\mathbb{E}[\Delta\mathbf{p}\Delta\mathbf{p}^{\mathsf T}]\ &\approx\mathbf{J}i\mathbf{R}{cw} \mathbb{E}[\Delta\mathbf{x}_w\Delta\mathbf{x}w^{\mathsf T}] \mathbf{R}{cw}^{\mathsf T}\mathbf{J}i^{\mathsf T}\ &=\mathbf{J}i\mathbf{R}{cw} \Sigma_i \mathbf{R}{cw}^{\mathsf T}\mathbf{J}_i^{\mathsf T}. \end{aligned} $$

原論文將它寫成

$$\Sigma'=\mathbf{J}\mathbf{W}\Sigma\mathbf{W}^\mathsf{T}\mathbf{J}^\mathsf{T}.$$

該式採用圍繞高斯中心的局部仿射近似，未精確涵蓋整個橢球的非線性透視變換。$x_c/z_c$ 的二階導數包含 $-1/z_c^2$ 與 $2x_c/z_c^3$，因此高斯相對深度越大、中心越靠近相機，忽略的二階項越明顯。光柵化器使用 near-plane guard band 排除極端高斯，背後對應的正是這個近似條件。

得到二維中心與二維協方差後，高斯在像素 $\mathbf{p}$ 上的 footprint 為

$$ g_i(\mathbf{p})= \exp!\left[-\frac{1}{2} (\mathbf{p}-\mu'_i)^\mathsf{T} (\Sigma_i^{\prime})^{-1} (\mathbf{p}-\mu'_i) \right], $$

其有效不透明度則是

$$ a_i(\mathbf{p})=\alpha_i g_i(\mathbf{p}). $$

「Splatting」即將一個三維高斯繪製成具有面積、方向和軟邊界的二維橢圓核，使其影響周圍一組像素。實現中不可能遍歷無限支撐域，因而會取馬氏距離閾值 $\tau$：

$$ D_i^2(\mathbf p) =(\mathbf p-\mu_i')^\mathsf T (\Sigma_i')^{-1} (\mathbf p-\mu_i') \le \tau^2. $$

這個橢圓在水平與垂直方向的最大半徑分別為

$$ r_x=\tau\sqrt{(\Sigma_i'){11}}, \qquad r_y=\tau\sqrt{(\Sigma_i'){22}}. $$

由 $[\mu'{i,x}-r_x,\mu'{i,x}+r_x]\times[\mu'{i,y}-r_y,\mu'{i,y}+r_y]$ 可快速得到軸對齊包圍盒，再判定它與哪些 tiles 相交。協方差投影公式到 tile 光柵化之間的連接，正是這個有限橢圓包圍盒。

Alpha 混合與像素合成

一個像素通常被多個高斯覆蓋。將它們按相機深度由近到遠排列後，像素顏色為

$$ \mathbf{C}(\mathbf{p})= \sum_{i=1}^{N} T_i(\mathbf{p}),a_i(\mathbf{p}),\mathbf{c}_i(\mathbf{d}), $$

其中

$$ T_i(\mathbf{p})= \prod_{j=1}^{i-1}\left[1-a_j(\mathbf{p})\right] $$

是光線到達第 $i$ 個高斯前仍未被遮擋的透射率。若考慮背景顏色 $\mathbf{c}_{bg}$，還可以補上

$$ T_{N+1}(\mathbf{p})\mathbf{c}_{bg}. $$

實際計算時不需要反覆展開乘積，可以使用前向遞推：

$$ \begin{aligned} \mathbf C_0&=\mathbf 0,\qquad T_1=1,\ \mathbf C_i&=\mathbf C_{i-1}+T_i a_i\mathbf c_i,\ T_{i+1}&=T_i(1-a_i). \end{aligned} $$

每處理一個高斯，只更新一次顏色與透射率。當 $T_i$ 足夠接近零，更遠的高斯對結果和梯度都幾乎沒有影響，提前終止（early termination）因而同時具有圖像形成與計算上的依據。

這個式子與 NeRF 離散體積渲染的圖像形成模型非常接近。差別在於，NeRF 通常從體密度 $\sigma_i$ 和採樣間隔 $\delta_i$ 計算

$$ a_i=1-\exp(-\sigma_i\delta_i), $$

3DGS 直接學習高斯的不透明度 $\alpha_i$，再乘上二維高斯 footprint。兩類方法都以「透射率 × 當前不透明度 × 顏色」累積像素，圖元與不透明度的來源則各不相同。

高斯核對二維中心可微：

$$ \frac{\partial g_i}{\partial\mu'_i} =g_i(\mathbf{p})(\Sigma_i^{\prime})^{-1} (\mathbf{p}-\mu'_i). $$

Alpha 混合對不透明度的梯度也能寫出很直觀的形式。令 $\mathbf C_{>k}$ 表示第 $k$ 個高斯後方所有內容（包括背景）在局部重新歸一化後的合成顏色，則整個像素可拆成

$$ \mathbf C =\mathbf C_{<k} +T_k\left[a_k\mathbf c_k+(1-a_k)\mathbf C_{>k}\right]. $$

對 $a_k$ 求導可得

$$ \frac{\partial\mathbf C}{\partial a_k} =T_k\left(\mathbf c_k-\mathbf C_{>k}\right). $$

這個式子比「整條鏈可微」更能說明訓練行為。前方圖元已接近不透明時，$T_k$ 很小，後方高斯很難獲得梯度；當高斯顏色與其後方合成色接近時，調整不透明度同樣不會顯著改變圖像。位置、顏色與不透明度因此存在互相補償的空間，僅靠圖像損失未必能得到唯一或幾何正確的場景。

圖像誤差仍可經過混合、二維 footprint、投影和協方差參數化，一路反向傳播至三維位置、尺度、旋轉、不透明度與 SH 係數。上面的梯度分析也預示了兩個常見問題：遮擋後方的幾何學得較慢，具有相似顏色的錯誤高斯則可能長時間留在場景中。

Tile-based 光柵化器與實時渲染

如果每個像素各自搜索所有高斯並排序，計算仍然會非常昂貴。3DGS 的 CUDA 光柵化器採用 tile-based 流程：

1. 將圖像平面劃分成 $16\times16$ 像素的 tiles。
2. 做視錐裁剪，只保留可能影響圖像的高斯；論文實現以高斯的 99% 置信區間判斷相交，並用 guard band 排除投影不穩定的極端高斯。
3. 計算每個二維高斯覆蓋哪些 tiles；若跨越多個 tile，就為每個相交 tile 生成一份索引。
4. 將 tile ID 與視空間深度組合成排序鍵，使用 GPU radix sort 一次完成分組與近似的深度排序。
5. 每個 tile 由一個線程塊（thread block）處理。高斯數據分批載入共享內存（shared memory），tile 內的像素執行各自的 Alpha 混合。
6. 當某個像素的不透明度已接近飽和，就停止處理更遠的高斯；當 tile 內所有像素都飽和，整個 block 可以提前結束。

令 $M$ 為視錐內高斯數，$K$ 為高斯與 tile 相交後生成的實例數，$P$ 為像素數，$\bar n$ 為提前終止前每個像素實際混合的平均高斯數。忽略固定次數的 radix-sort passes，主要工作量可以概括為

$$ \underbrace{O(M)}{\text{投影與裁剪}} +\underbrace{O(K)}{\text{複製、排序與分組}} +\underbrace{O(P\bar n)}_{\text{像素混合}}. $$

這個估算說明渲染時間並非只由高斯總數決定。大 footprint 會同時推高 $K$ 和 $\bar n$，深度重疊嚴重的場景也會增加 $\bar n$；反過來，即使模型包含很多高斯，只要大多數高斯很小、可快速裁剪，仍可能維持較高幀率。評估壓縮或剪枝方法時，僅報告高斯數量容易漏掉真正的光柵化負擔。

tile 層級排序會讓某些高斯交疊情況下的 Alpha 混合帶有近似性。原論文指出，當訓練收斂且 splat 接近像素尺度時，這類近似通常不會形成可見偽影。統一的 tile 列表同時帶來規則的內存訪問和大規模並行計算。

反向傳播會重用前向已排序的 tile 列表，從後向前遍歷並恢復所需的中間量，免去為每個像素保存任意長度混合歷史的開銷。3DGS 設計中容易被公式掩蓋的一點是，實時性來自表示、近似排序、內存布局與停止條件共同配合，單獨把高斯投影公式搬到 GPU 並不會自然得到同樣的速度。

渲染質量、速度與存儲開銷

原論文在 Mip-NeRF 360 數據集上同時報告圖像指標、訓練時間、渲染幀率與模型存儲 [@kerbl2023gaussians]：

$7\mathrm{k}$ 到 $30\mathrm{k}$ 迭代之間，PSNR 提高 $1.61,\mathrm{dB}$，LPIPS 從 $0.279$ 降到 $0.214$，模型存儲則從 $523$ MB 增加到 $734$ MB，幀率從 $160$ 降到 $134$。訓練後期增加的圖元和外觀容量改善了質量，也同時增加投影、排序和存儲成本。

Mip-NeRF 360 的三項質量數值由 3DGS 論文從原工作轉錄，效率數據也受實現與硬件影響，因此這張表更適合用來讀取數量級和方法內部的質量—成本曲線。若要比較當代實現，還需要統一 GPU、分辨率、渲染代碼與計時邊界。

高斯集合的優化與密度控制

3DGS 在更新高斯連續參數的同時，也會離散地調整高斯集合。兩類操作交錯進行，大致可以寫成：

flowchart TD
    A["SfM 相機與稀疏點雲"] --> B["初始化高斯位置、尺度、不透明度與顏色"]
    B --> C["抽取訓練視角"]
    C --> D["投影與 tile 光柵化"]
    D --> E["合成圖像"]
    E --> F["計算重建損失"]
    F --> G["反向傳播並更新高斯參數"]
    G -->|"下一次迭代"| C
    G -.->|"每隔若干迭代"| H["Clone / Split / Prune / Opacity Reset"]
    H --> C

初始化

3DGS 從 SfM 稀疏點的位置開始。每個初始高斯先設為各向同性，其尺度參考最近三個 SfM 點的平均距離。這提供了比完全隨機分布更可靠的幾何起點，但初始點仍遠不足以覆蓋完整表面，因此必須在訓練中增密。

Clone：欠重建區域的複製

論文以視空間位置梯度的平均幅值判斷哪些高斯仍在努力移動以降低誤差。若第 $i$ 個高斯在 $M_i$ 次可見觀測中的二維中心為 $\mu'_{i,m}$，可以把這個統計量概括成

$$ g_i =\frac{1}{M_i} \sum_{m=1}^{M_i} \left| \frac{\partial\mathcal L_m}{\partial\mu'_{i,m}} \right|_2. $$

若 $g_i$ 超過閾值，而且高斯尺度較小，通常表示現有圖元太少、無法覆蓋附近結構。此時複製一個同尺度高斯，並沿位置梯度方向移動副本。

Clone 同時增加高斯數量與局部覆蓋體積，適合補上欠重建區域。不過 $g_i$ 是屏幕空間信號，會受到焦距、圖像分辨率、相機距離與可見視角數影響。固定的 $0.0002$ 閾值具有明顯的數據與實現尺度，難以視為場景幾何誤差的普適界線。

Split：大尺度高斯的分裂

位置梯度與高斯尺度同時偏大，通常說明單一圖元負責的範圍過廣。3DGS 以兩個較小高斯取代它：新中心從原高斯分布中採樣，尺度除以實驗設定的 $1.6$。

Split 主要提高局部表示分辨率，使一個模糊的大橢球逐漸分工成數個能貼近細節的小橢球。這裡可以檢查「分裂後體積延續」的含義。若以一個協方差橢球的尺度體積

$$ V_i\propto\sqrt{\det\Sigma_i} =s_{i,x}s_{i,y}s_{i,z} $$

衡量，兩個 child 的三軸尺度都除以 $1.6$ 時，總尺度體積與 parent 的比值為

$$ \frac{V_{\mathrm{children}}}{V_{\mathrm{parent}}} =\frac{2}{1.6^3} \approx0.488. $$

所以論文所描述的體積延續，更適合解讀為兩個新中心對原空間覆蓋範圍的延續，協方差橢球體積並未嚴格守恆。$1.6$ 是經驗設定，它同時增加基底數量並收縮單個基底的支撐域。

不透明度也有類似問題。若把一個有效不透明度 $a_p$ 的 parent 換成 $n$ 個完全重合、顏色相同的 children，每個 child 的不透明度為 $a_c$，合成後為

$$ a_{\mathrm{eff}}=1-(1-a_c)^n. $$

要求分裂前後立即保持相同遮擋量，應取

$$ a_c=1-(1-a_p)^{1/n}. $$

3DGS 的密度控制沒有把這類守恆式設為硬約束，分裂後仍交給圖像損失、不透明度重置與剪枝共同修正。從這個角度看，clone 與 split 更接近模型結構的啟發式搜索，並不保持某個連續場完全等價的解析細分。

Prune：剪枝與不透明度重置

持續增密會令高斯數量失控，也會留下漂浮點。3DGS 會移除不透明度低於閾值、世界空間尺度過大或屏幕 footprint 過大的高斯；訓練中還會周期性把不透明度重置到接近零，讓有穩定圖像貢獻的高斯重新獲得較高不透明度，其餘高斯則在後續剪枝中消失。

原論文的參考設定包括：預熱後每 100 次迭代做一次增密，以平均視空間位置梯度 $0.0002$ 作為候選閾值，並每 3000 次迭代重置不透明度。這些數字屬於論文實驗的超參數，會隨數據尺度、圖像分辨率與方法改動而調整。

3DGS 的完整管線

上述機制可分成三個彼此咬合的循環：

令 $N$ 為當前高斯數量，$\Theta_N$ 為全部連續參數，訓練目標可以抽象成

$$ \min_{N,\Theta_N} \sum_{m} \mathcal L!\left( \mathcal R(\Theta_N;\mathcal C_m), \mathbf I_m \right), $$

其中 $\mathcal R$ 是可微光柵化器，$\mathcal C_m$ 是第 $m$ 個相機。Adam 在固定 $N$ 時更新 $\Theta_N$；clone、split 與 prune 則改變 $N$ 和參數空間本身。這是一個混合了連續參數與離散結構的優化問題，原方法以交替啟發式操作近似求解。

更合適的理解是把一組高斯看成可增刪的局部基底。位置、尺度、旋轉、SH 與不透明度是基底的係數和形狀，圖像梯度衡量現有基底能否解釋觀測，增密則在殘差較大的位置增加自由度。Prune 和 opacity reset 承擔近似的複雜度控制，儘管目標函數裡沒有顯式寫出 $\beta N$ 這類模型規模懲罰。這個視角也能解釋為何單純沿用高斯表示、卻拿掉密度控制，通常很難重現完整 3DGS 的效果。

3DGS 的表示範圍

場景表示與物體表面

3DGS 的直接目標是新視角圖像質量，輸出為一組體積式高斯圖元，沒有三角網格所具備的拓撲連接關係。高斯中心和扁平橢球經常落在表面附近；可測量表面、法線或網格仍需要額外幾何約束或表面提取。

高斯的方向與外觀屬性

每個高斯除了位置，還有完整的方向性尺度、不透明度和視角相關顏色；投影後的二維 footprint 也會隨相機與深度改變。「帶半徑的點」不足以描述協方差投影與 Alpha 混合所帶來的行為。

與體積渲染的關係

3DGS 與體積渲染共享相近的 Alpha 合成形式。它直接光柵化顯式高斯圖元，沒有沿光線積分連續密度，渲染中還包含局部投影與 tile 級排序近似。

增密機制的判據

3DGS 根據視空間位置梯度選取候選，再由尺度區分 clone 與 split，並以剪枝和不透明度重置抑制冗餘。這套判據分別處理欠重建、過度粗糙的表示與無效圖元。

限制與適用範圍

3DGS 具備顯式結構、快速訓練、實時高分辨率渲染，以及比普通點雲更連續靈活的表示，同時受到以下條件限制 [@chen2024survey]：

• 它依賴已標定的多視角圖像與足夠可靠的 SfM 初始化；稀疏視角或錯誤位姿會直接影響結果。
• 優化目標以圖像重建為主，幾何可以被外觀補償，因而不保證表面精確。
• SH 能擬合視角相關顏色，但不會自然分離材質、光照、陰影與幾何誤差。
• 數十萬乃至數百萬個高斯會帶來存儲與顯存成本。
• 大型、近相機或交疊複雜的高斯會放大投影與排序近似造成的偽影。
• 3DGS 面向靜態場景，沒有直接處理物體運動、形變或時間一致性。

後續 3DGS 研究由此延伸出抗混疊、幾何正則、壓縮、大場景、稀疏輸入、動態建模、材質與光照解耦等方向。

結語

3DGS 最值得延續的思想，是讓場景基底、可微渲染與 GPU 執行模型彼此配合。協方差給局部基底以方向和尺度，投影雅可比把三維擾動傳到像素，Alpha 混合建立遮擋與梯度，tile 光柵化器再把這些計算整理成 GPU 能高效執行的數據流。高斯集合還能在訓練中改變規模，表示容量由殘差逐步長出來。

這條管線也留下了清楚的近似邊界：透視投影只在中心附近線性化，tile 共用排序不能涵蓋所有逐像素深度關係，圖像損失允許外觀補償幾何，clone 與 split 又依賴視空間閾值和經驗倍率。後續研究中的抗混疊、幾何正則、重新參數化與更穩定的密度控制，都可以追溯到其中某一處近似。

理解 3DGS 時，還需要追問每個公式向實現做了哪些讓步：協方差如何變成有限包圍盒，透射率何時讓梯度消失，分裂是否真的守恆，高斯數量又如何影響 tile 實例數。沿著這些問題往下走，後續方法的改動位置與代價會比單純記住一張流程圖清晰得多。

@article{kerbl2023gaussians,
  author = {Kerbl, Bernhard and Kopanas, Georgios and Leimkühler, Thomas and Drettakis, George},
  title = {3D Gaussian Splatting for Real-Time Radiance Field Rendering},
  journal = {ACM Transactions on Graphics},
  volume = {42},
  number = {4},
  pages = {139:1--139:14},
  year = {2023}
}

@article{chen2024survey,
  author = {Chen, Guikun and Wang, Wenguan},
  title = {A survey on 3d gaussian splatting},
  journal = {ACM Computing Surveys},
  year = {2024}
}

To avoid confusion, it is important to note that Valve has used the “Steam Machine” label before. In the mid-2010s, Valve partnered with OEMs such as Alienware, Zotac, and others to ship small form factor PCs that carried the SteamOS branding or were marketed as “Steam Machines.” Those devices did reach the market, but they were essentially prebuilt PCs: each OEM defined its own chassis, power envelope, and component mix, and the whole initiative quietly faded without ever establishing a coherent console platform. This article is not about that first wave of OEM Steam Machines. It examines the Steam Machine that Valve announced in November 2025: a unified, Valve-designed living-room box that, at the time of writing, still exists only as an announced product, not as something you can actually buy.

Even before launch, part of the 2025 Steam Machine’s likely struggle can be read from its identity. It is attempting to absorb the ease and predictability of consoles while carrying the complexity and variability of PC hardware. It promises openness but, at least at this stage, lacks the guaranteed audience that would give studios a clear incentive to optimize for it. And it is trying to appeal to the living-room user without yet offering the frictionless coherence that defines every successful console generation. In this unresolved middle ground, the Steam Machine risks being too open in philosophy to function as a traditional console, too constrained in positioning to satisfy PC purists, and too undefined to court the mainstream.

The proposed hardware specification itself tells an interesting story. Valve’s modernized Steam Machine reference configuration is described as featuring a Zen 4 six-core processor capable of boosting to roughly 4.8 GHz, paired with a semi-custom RDNA 3 GPU with 28 compute units and 8 GB of dedicated GDDR6 VRAM. This is said to sit alongside 16 GB of DDR5 system memory, NVMe SSD storage options, DisplayPort 1.4 output capable of 4K/240 Hz or even 8K/120 Hz, and a chassis that measures barely more than a six-inch cube. On paper, this is not a low-end device; it is closer to a compact gaming PC dressed as a console. But crucially, all of this still lives in spec sheets, decks, and press briefings — not in shipping hardware.

For clarity, here is a concise comparison between the projected Steam Machine configuration and the two major ninth-generation consoles. This is a comparison of targets and claimed specifications, not of devices that currently coexist on retail shelves:

On paper, it is an impressive lineup, even daring in places. But hardware alone never defines a platform — and projected hardware, even less so. Consoles thrive because they are unified systems: hardware, operating system, developer tools, online services, and business model all operate in lockstep. The earlier OEM Steam Machine program failed precisely because Valve provided the software layer and the brand, but outsourced the hardware story to “dozens of OEMs” that produced mutually incompatible designs, each with different thermals, acoustics, and price points. Developers were expected to treat that mess as a single “platform,” despite the lack of a fixed hardware baseline.

The 2025 Steam Machine is, at least in theory, an attempt to correct that error. Valve now presents a single reference box with a tightly defined specification and a clear industrial design, rather than a scattershot catalog of partner machines. Yet even in this revised form, the structural risks have not disappeared. Valve has shown no sign that it intends to follow the traditional console strategy of heavily subsidizing hardware to drive long-term software revenue. If Valve expects manufacturing and distribution to be handled on commercial terms, retail prices are likely to end up much closer to high-end small-form-factor PCs than to mass-market consoles.

At the same time, SteamOS as a living-room platform is still young. Proton continues to improve, but it is not perfect, and the library of native Linux titles remains relatively small. A device that would depend on game sales to drive hardware adoption and hardware adoption to justify game development risks stalling if either side hesitates. The original OEM Steam Machines never escaped that loop; the 2025 Steam Machine has not yet demonstrated, beyond marketing language, how it will.

Viewed through the broader lens of gaming history, the Steam Machine belongs to a familiar lineage of “in-between” hardware — devices like the 3DO, Apple Pippin, and Ouya — built around the hope that console convenience and PC openness can be reconciled in one product. History suggests that the space between those worlds is unstable. Consoles draw strength from coherence; PCs from flexibility; hybrids must somehow preserve both without inheriting the limitations of each. The OEM Steam Machines of the last decade failed that test, and the 2025 machine has not yet shown how it will answer it.

Ironically, Valve has already demonstrated that it understands this divide. The Steam Deck has succeeded precisely because it embraces the opposite philosophy: strict hardware unification, a tightly controlled OS stack, deep Proton integration, and a user experience designed around a single chassis rather than a nebulous family of boxes. In many ways, the Deck feels like the Steam Machine idea re-imagined with the discipline it always needed — discipline that the OEM program never had, and that the 2025 Steam Machine will have to prove it can sustain once it moves beyond trailers into reality.

The Steam Machine is therefore best understood as an unresolved experiment rather than a failed product. It asks whether the openness of the PC can survive the coordination demanded by the living room, and whether Valve is willing to impose enough discipline to make that answer convincing. Until the hardware moves beyond specifications and stage demonstrations, the machine remains a glimpse of a parallel living-room future — visible, carefully drawn, and still just out of reach.

這是一個真正符合自己興趣的課題，所以我決定主動出擊，而不是等待系統隨機分配。

一、為何選擇主動申請

過往參與的多個嵌入式及車輛工程相關項目，讓我在軟硬件協作、數據流程及視覺模型測試方面累積了一定經驗。
導師給出的其中一個課題，在數個模組上都能與我的背景自然對應。

主動申請並不是「爭位置」，而是讓導師看到：這個方向對我來說不是臨時起意，而是與過去經驗及未來規劃同時對齊的選擇。

二、申請信的思路：重點在於「對齊」，而非「表現」

雖然只是一封寫給導師的電郵，但我以「工程師向項目負責人申請加入」的方式來撰寫內容，目的只有一個：
讓導師能在短時間內判斷我是否適合這個課題。

整體結構包括：

1. 清晰列出自己曾處理過、而且與課題真正相關的部分
   不誇大，不堆砌，只展示具體能力邊界。

2. 說明目前正在學習或補強的方向
   讓導師看到持續投入，而非臨急抱佛腳。

3. 表達真實動機，但保持克制
   不使用空洞語氣，也不作無法預估的承諾。

4. 強調會自行準備，不額外增加導師負擔
   減少導師在心理上的不確定成本，使選擇更容易。

整封電郵保持專業、節制、重點明確，強調的是「能力匹配」及「方向一致」。

::: details, title="脫敏後電郵"

【老師A】您好：

我是《【課程A】》課程的學生【姓名A】。得知您公布的畢業設計課題，其中 「【課題A】」 與我的技術基礎及未來規劃高度契合，因此特向您表達真誠的意願，希望能申請參與該課題，並在正式開始前提前做好準備工作。

在課程項目中，我主導完成了《【項目A】》的整體架構設計，包括 【技術A1】多任務調度、【技術A2】多模組協同 以及 【技術A3】多來源資料整合，實現了 【功能A1】、【功能A2】與【功能A3】 等能力。這項工作讓我對系統協作、實時處理及資料鏈路有了完整工程上的理解。

在【領域A】方面，我曾基於 【框架A】 完成相關模型訓練，包括 【模型A】 與 【模型B】 的實驗，內容涉及 【技術B1】、【技術B2】、【技術B3】與【技術B4】。近期亦在 【平台A】 上進行開發測試，對端側平台的部署流程及效能限制有一定了解。目前正系統學習 【框架B】、【框架C】與【框架D】 的模型轉換及加速方法，希望在後續課題中探索 【技術C1】—【技術C2】融合方案 的效率優化及部署策略。

了解到您團隊在 【研究方向A】 方面具有長期積累，我非常希望能在您的指導下參與此課題，並會提前閱讀相關資料、補強必要背景。如您方便，也希望能向您請教課題初期的準備重點，以便更有方向地開展前期工作。

感謝您在百忙中閱讀此信，期待您的回覆。

此致 敬禮

:::

三、導師的回覆：直接而肯定

在發出申請後不久，我收到導師簡潔而明確的回覆：

「那我就將這個機會留給你，XXXXXXXXXXXXXXXX，加油。」

這代表三件事：

• 課題已正式交由我負責
• 之後或會有更深入的協作環節
• 導師對我的準備及動機給予肯定

對畢業設計來說，這是一個相當理想的開始。

四、接下來的推進方式

在項目初期，我傾向以工程化的節奏前進：
先讓整個系統順利運作，理解資料流程，再按實際需求逐步展開相關實驗。

與其急於展示某個「亮眼的成果」，更重要的是真正理解整個系統的協作方式，這亦較符合系統類課題的自然節奏。

五、總結：不是「套瓷」，而是準備在正確時機對上號

不少人會把這種行為戲稱為「套瓷」。
但真正做過之後便會發現：當你對自己的方向足夠明確、在相關領域已有實際積累，並在關鍵時刻主動表達，結果往往比想像中順利。

這次成功爭取課題，並不是甚麼「技巧性的故事」，而是準備與機會在正確時機相遇。

未來的項目會按導師安排逐步展開，而我也會在過程中繼續補強自己對系統、數據流程及工程實務的理解。
這不是甚麼華麗的開始，但卻是一個踏實可靠的起點。

Spoiler

This post contains ||**spoiler content**|| which only appears when you hover or tap.

This post contains ||spoiler content|| which only appears when you hover or tap.

Underline emphasis

Use ++keywords++ to highlight certain keywords.

Use ++keywords++ to highlight certain keywords.

Highlighter emphasis

Use ::key terms:: to mark important wording.

Use ::key terms:: to mark important wording.

Details

::: details, title="Tap to reveal more" This block accepts regular Markdown, such as lists:

• Point one
• Point two

Images, links, and any other inline content work as well. :::

Banner

Info

::: banner{info}
This is the Info style.
:::

::: banner{info} This is the Info style. :::

Success

::: banner{success}
Completion notice.
:::

::: banner{success} Completion notice. :::

Warning

::: banner{warning}
This is a **banner** reminder block.
:::

::: banner{warning} This is a banner reminder block. :::

Error

::: banner{error}
Error message.
:::

::: banner{error} Error message. :::

Grid

Image Mode

::: grid {cols=2,gap=1.2rem,type=images}
![Image 1](https://img.nishikori.dev/2025/11/20251109215205561.jpeg)

![Image 2](https://img.nishikori.dev/2025/11/20251108212225527.jpeg)
:::

::: grid {cols=2,gap=1.2rem,type=images}

:::

Mixed Mode

::: grid {cols=2,gap=1.2rem}
![NGC](https://img.nishikori.dev/2025/11/20251109215205561.jpeg)

![PS3](https://img.nishikori.dev/2025/11/20251108212108445.jpeg)

Left column content

Right column content
:::

::: grid {cols=2}

Left column content

Right column content :::

Video Embed

YouTube

https://www.youtube.com/watch?v=dX8gZubnMjA

https://www.youtube.com/watch?v=dX8gZubnMjA

Bilibili

https://www.bilibili.com/video/BV1xx411c79H

https://www.bilibili.com/video/BV1xx411c79H

Spoiler

本文會提到||**劇透內容**||，只有指標移上去或點一下才會顯示。

本文會提到||劇透內容||，只有指標移上去或點一下才會顯示。

新增標記語法

下劃線強調

使用++關鍵詞++突出某些關鍵詞。

使用++關鍵詞++突出某些關鍵詞。

螢光筆強調

使用 ::重點詞彙:: 來螢光標記某些重點詞彙。

使用 ::重點詞彙:: 來螢光標記某些重點詞彙。

Details

::: details, title="點我展開更多內容" 這段文字可以寫 Markdown，例如列表：

• 第一點
• 第二點

也可以放圖片、連結等等。 :::

Banner

Info

::: banner{info}
這是 Info 風格。
:::

::: banner{info} 這是 Info 風格。 :::

Success

::: banner{success}
完成提示。
:::

::: banner{success} 完成提示。 :::

Warning

::: banner{warning}
這是一個 **banner** 提醒區塊。
:::

::: banner{warning} 這是一個 banner 提醒區塊。 :::

Error

::: banner{error}
錯誤訊息。
:::

::: banner{error} 錯誤訊息。 :::

Grid

圖片模式

::: grid {cols=2,gap=1.2rem,type=images}
![圖片一](https://img.nishikori.dev/2025/11/20251109215205561.jpeg)

![圖片二](https://img.nishikori.dev/2025/11/20251108212225527.jpeg)
:::

::: grid {cols=2,gap=1.2rem,type=images}

:::

混合模式

::: grid {cols=2,gap=1.2rem}
![NGC](https://img.nishikori.dev/2025/11/20251109215205561.jpeg)

![PS3](https://img.nishikori.dev/2025/11/20251108212108445.jpeg)

左邊的內容

右邊的內容
:::

::: grid {cols=2}

左邊的內容

右邊的內容 :::

Video Embed

YouTube

https://www.youtube.com/watch?v=dX8gZubnMjA

https://www.youtube.com/watch?v=dX8gZubnMjA

Bilibili

https://www.bilibili.com/video/BV1xx411c79H

https://www.bilibili.com/video/BV1xx411c79H

From the heavy SCPH-1000 to the featherlight PS one, every version carries its own texture, sound, and character.
For collectors, choosing your PlayStation is no longer only about playing games. It is also about choosing which moment of the 1990s you want to hold.

The First Form — SCPH-1000

Released in Japan in December 1994, the SCPH-1000 still bears the marks of a launch machine. It offered an unusually generous array of connections — S-Video, RCA, Parallel I/O, Serial I/O, and the soon-to-be-standard AV Multi Out. Inside, its early PU-7 motherboard and audio hardware helped give the model a second reputation: that of an unexpectedly capable standalone CD player.

But it ran hot, aged quickly, and its laser assemblies often failed.
When Sony prepared for international launch, the quiet revisions began.

The Global Debut — SCPH-1001 / 1002 / 3000 / 3500

The export models dropped the S-Video port and switched to the more reliable PU-8 board.
Reliability improved, but the console still struggled with FMV skipping and heat buildup.
The SCPH-3000 and 3500 streamlined production while keeping most expansion ports intact.
Heavy, solid, unmistakably mid-’90s — these are the machines in which Sony’s first idea of the PlayStation began to settle into a durable form.

The Maturity — SCPH-500x / 550x

By 1996, the PlayStation had grown up.
The 550x series introduced a digital servo optical drive, repositioned the disc assembly to reduce heat, and removed the rear RCA jacks in favor of a single AV Multi Out — the connector still used for RGB output today.
For many owners, it is the most convincing midpoint: stable, easy to modify, and still close to the original vision.

The Refinement — SCPH-700x / 750x / 900x

Late-era PlayStations focused on integration and efficiency.
The 750x featured a sharper video encoder and a sturdier metal-beam drive, while the 900x — the “final classic shell” — eliminated the Parallel I/O port entirely.
That move ended the era of cheat cartridges and hardware debugging tools, but created one of the most reliable PS1s ever made.

If you simply want to play games, these revisions are among the least troublesome to own.

The Reinvention — PS one (SCPH-10x)

In 2000, Sony reimagined the PlayStation as the PS one — smaller, brighter, almost friendly.
It moved the power supply outside, leaving only two rear ports: power and AV Multi Out.
Internally, the compact PM-41 board ran cooler and quieter, though at the cost of all expansion interfaces.

This wasn’t a downgrade — it was a statement.
The PS one embodied Sony’s late-era minimalism: white, soft, and approachable.
It blurred the line between hardware and home décor.

Which PlayStation Is Yours?

There’s no single “best” PS1.
Only the one that fits who you are.

• For collectors:
  Hunt down the SCPH-1000 or early 1001/1002 units.
  They’re raw, imperfect, and carry the aura of the launch era.
  Expect to repair lasers or capacitors — that’s part of their charm.

• For practical players:
  The 550x–900x range is the sweet spot.
  Reliable drives, great compatibility, and easy maintenance.
  A clean SCPH-900x is a wonderfully undramatic machine to own.

• For minimalists:
  Go with the PS one.
  Small, quiet, perfectly compatible, and beautiful on a desk.
  You lose expansion ports, but gain simplicity and silence.

The Sound of Continuity

Owning a PlayStation isn’t just about specs.
It’s about time — the story of Sony learning how to build a console that defined a decade.
Each revision is a chapter in that journey, from over-engineering to refinement.

Whichever model you choose, that startup chime — half mechanical, half dreamlike — remains unchanged.
The boards beneath it grew smaller, the ports disappeared, and the shell eventually turned white; the sound still arrives as though no time has passed at all.

The evolution of game controllers isn’t just a story of technology;
it’s a story about how humans learned to speak to machines.

From Obedience to Expression

Returning to the Atari 2600, I was struck by how much of both hands the controller demanded. One hand moved the joystick while the other steadied the base and reached for its single button. The arrangement was direct and physical, but it left little room for actions to overlap. The player learned the shape of the machine before the machine could disappear into play.

Then came Gunpei Yokoi’s D-pad on the Nintendo Entertainment System (NES).
With one thumb handling all directions, the other fingers were freed —
you could move and jump, attack, or use items simultaneously.
What looked like a tiny cross of plastic turned out to be a quiet revolution.

The D-pad did not invent direction or action, but it separated them with unusual clarity. That small act of compression created a design space where combinations felt natural:

• Move + Jump (Super Mario Bros.)
• Move + Shoot (Contra)
• Navigate Menu + Use Items (The Legend of Zelda)

Once movement belonged to one thumb, designers could begin composing around what the rest of the hand was free to do.

The Birth of Combinational Input

This “combinational input” was the seed of modern action-game complexity.
It pushed designers to think not in isolated button presses, but in relationships —
between movement and rhythm, physics and timing, systems and feedback.
From this point onward, the controller could recede slightly, leaving more room for timing, combination, and expression.

The contrast is not as simple as one era obeying machines and another escaping them. What Yokoi understood was subtler: the thumb’s motion is compact and almost instinctive, while digital games require discrete, readable signals. The D-pad placed those two facts in the same piece of plastic and let players build a language from the result.

The Interface Becomes Human

Behind this was a quieter change in interface philosophy. A good controller did not need to announce its cleverness; it could withdraw beneath the hand until intention and action seemed almost continuous. The console remained an electronic box, but the distance between that box and the player’s body had begun to narrow.

A Continuing Evolution

After the D-pad, the story continued:

• 1996 — N64’s analog stick enabled continuous 3D motion
• 1997–98 — DualShock defined dual-stick precision
• 2006 — Wii Remote redefined physical “action”
• 2017+ — Joy-Con, Adaptive Triggers, touchpads expanded sensory dimensions

Yet all these advances trace back to one simple question:

“Can players feel freer — using only one thumb?”

Gunpei Yokoi’s answer was a cross of plastic small enough to disappear beneath the thumb. Its achievement was not that it made the controller expressive on its own, but that it left more of the expression to the player.

Although I previously categorized the Switch 2 as a tenth-generation console, its lifespan overlaps extensively with the PS5 and Xbox Series era. Its ecosystem remains continuous with the original Switch, and much of its early development pipeline is still tied to Switch 1 practices. For the purposes of this essay, I therefore consider it part of the ninth-generation landscape.

:::

From the end of 2020 to the close of 2025, the contours of this generation gradually became easier to read. Five years exposed the consequences of its hardware compromises and changes in platform strategy. If the old rules of the console race still mean anything — performance, third-party support, flagship software, and a coherent hardware identity — then the PlayStation 5 is the clearest winner of the generation. It is not flawless; it is simply the machine that remained closest to those rules.

Xbox Series reoriented itself toward services; Nintendo released a machine that thrives commercially but fails technically as a next-generation device. Against this shifting background, the PS5—despite its own shortcomings—remained the only console still running the “classic” race: performance, third-party alignment, flagship titles, and coherent platform identity.

The Performance Leap: A “Next-Gen” Change Players Could Feel

The PS5’s success begins with its decisive escape from the PS4’s deepest structural weakness. The PS4 era was shackled by the underpowered Jaguar CPU, which constrained every domain from world density and AI complexity to frame pacing and streaming logic. The PS5’s leap to Zen 2 transforms the entire performance landscape: wide memory bandwidth, a modern GPU pipeline, and a deeply customized SSD architecture elevate everything from asset throughput to open-world traversal.

The Xbox Series X is formidable on paper, but Microsoft’s dual-SKU strategy complicated the story. Supporting the weaker Series S established a lower common baseline, and the Series X’s raw power did not always translate into advantages that players could readily see.

Meanwhile, Nintendo’s Switch 2 offers a spectacular jump relative to its predecessor, but its hybrid design places intrinsic limits on performance in handheld mode. Heat dissipation, power management, and battery constraints collide, preventing its impressive theoretical performance from manifesting consistently in real gameplay.

Of the three, the PS5 offered the clearest version of a generational shift that could be felt not only in specifications, but on screen.

Pricing and Platform Value: Sony’s Commitment to the Classic Console Model

Sony’s pricing strategy hews closely to the decades-tested logic of the console business: a stable $399/$499 structure, adjusted regionally through promotions rather than fundamental restructuring. This maintains a clear psychological entry point for new players.

Microsoft’s Series X, pushed upward toward 599 USD in certain regions, weakens the value argument it once held. The Series S remains inexpensive, but places visible constraints on the shared platform baseline. At 449 dollars, the Switch 2 occupies a different and sometimes awkward category: priced beyond the usual expectations of a handheld, yet not designed to compete directly with high-end living-room hardware.

In this environment, Sony’s approach—traditional, predictable, and steady—stands in stark contrast to the experimental or compromised strategies of its competitors.

Market Scale: From a Historic Rivalry to an Era of Structural Imbalance

In the Xbox 360 and PS3 era, the battlefield was tight, messy, and fiercely competitive.
In the ninth generation, the gap is structural.

PS5 maintains a substantial lead over Xbox Series, and the direction of travel shows little sign of reversing. Switch 2, despite a remarkably fast launch, occupies a different category altogether — often a companion to high-performance hardware rather than its direct replacement.

Sony wins this era not through dominance of a crowded field, but through persistence in a field that others abandoned.

Game Output: A Three-Tier Structure That Defines the Generation

The PS5’s software ecosystem forms a stratified, resilient structure. Technical showcases such as Demon’s Souls, Ratchet & Clank: Rift Apart, and Returnal demonstrate the console’s architectural edge.
Blockbuster centerpieces—Marvel’s Spider-Man 2, God of War Ragnarök, Final Fantasy XVI, Final Fantasy VII Rebirth, Death Stranding 2—consolidate Sony’s long-standing cultural capital.

And in the mid-budget space, where creativity and identity often flourish, the PS5 saw releases such as Ghost of Yōtei, which drew attention through its atmosphere, combat sensibility, and artistic voice. The breadth of this ecosystem became one of Sony’s clearest advantages.

Xbox’s first-party cadence lagged, and once-iconic series like Halo struggled to define themselves in the new era. Nintendo delivered the expected brilliance of its first-party titles, but could not meaningfully expand its third-party presence due to constraints inherent in its hardware.

The Structural Failure of the Nintendo Switch 2: A Commercial Success, but Not a Next-Gen Console

Switch 2 will surely be commercially successful. It already is. But commercial momentum and hardware success are not synonymous. As a next-generation system, the Switch 2 contains a series of structural flaws—technical, ergonomic, and ecosystem-level—that prevent it from standing on equal footing with the PS5 or even fully expressing its own theoretical capabilities.

Its shortcomings can be understood across three intertwined dimensions: performance constraints, visual presentation limits, and ecosystem stagnation.

Performance Trap: The Inescapable Triangle of Heat, Power, and Battery Life

The Switch 2’s hybrid nature forces mutually incompatible requirements onto its architecture.
In handheld mode, heat and battery constraints necessitate aggressive downclocking. In warmer climates, the system may reduce performance to avoid thermal faults. These design realities cripple its ability to deliver stable high-end rendering—even if the underlying silicon is capable of far more.

It is a device with strong theoretical potential but fundamentally constrained real-world performance.
By the very nature of its design, it cannot run in the same race as performance-driven living-room systems.

Visual Ceiling: A Screen That Cannot Reflect the Console’s Generational Leap

The Switch 2’s internal performance upgrade is not matched by its display. A screen with limited luminance, modest refresh behavior, and conservative resolution cap prevents the console’s visual advancements from being fully perceived.

The generational leap exists inside the silicon, but not on the surface where players experience it.
This disconnect undercuts the core emotional payoff of a next-gen transition.

Development Stall: A Next-Gen Console Without Next-Gen Development Cycles

The most damaging flaw comes from the development environment.
In the crucial early months of the platform’s life, access to Switch 2 dev kits was tightly restricted, forcing many studios to continue building games around Switch 1 baselines. Enhancements for Switch 2 had to be layered via compatibility patches rather than designed into the architecture from the start.

The result was a hesitant next-generation identity during the platform’s formative period. Many developers could not yet redesign world structure, material budgets, or simulation complexity around the new baseline, so players often encountered “upgraded Switch 1” games before distinctly Switch 2 experiences had time to emerge.

Switch 2’s early success therefore says at least as much about Nintendo’s first-party ecosystem and the continuity of the Switch audience as it does about the new hardware itself.

The Xbox Series Dilemma: A Dual-SKU Trap, Ecosystem Drift, and Microsoft’s Strategic Exit

If the Switch 2’s flaws arise from its hybrid ambition, the Xbox Series falters for entirely different reasons: product strategy misalignment, ecosystem inconsistency, and Microsoft’s deliberate pivot away from hardware-centric competition.

The Dual-SKU Trap: How Series S Became the Platform’s Lowest Common Denominator

Launching the Series X and Series S together was meant to widen the market. Instead, it constrained the entire ecosystem.
Microsoft mandates that all Xbox titles must support both machines. But the Series S has significantly less memory, lower GPU bandwidth, and stricter rendering budgets. This low baseline dictates the design of every multiplatform release.

The consequences have become increasingly visible. Black Myth: Wukong faced severe optimization challenges on Xbox, leading to a delayed release months behind the PS5 and PC versions. Developers of Battlefield 6 publicly stated that certain levels crashed entirely on Series S due to memory shortages.

The Series X is powerful—but the Series S becomes its ceiling.

Development Ecosystem Weakness: A Generation Without Strong Narrative Anchors

Technical issues compound systemic ones. Microsoft acknowledged performance problems in certain “optimized” titles and worked with partners to address them, revealing the fragility of its optimization pipeline. This stands in sharp contrast to the relative stability of PS5 builds.

First-party output, once the soul of Xbox, lost momentum. Halo, Gears, and Forza no longer anchor the generation the way they did in the past, and new IP struggled to fill the vacuum. Without generational milestones, the Xbox Series never shaped a coherent identity.

Microsoft’s Strategic Pivot: From Console Competitor to Platform Operator

The most consequential change to Xbox’s generational position is its own strategic evolution. Microsoft no longer positions Xbox as a hardware-centric business. Game Pass, xCloud, unified GDK pipelines, and multi-platform releases—even on PlayStation and Switch—signal a future where Xbox is a service, not a console.

This direction is commercially sound but tactically incompatible with “winning a console generation.”
The Xbox Series becomes one of many endpoints, not a flagship device.

Microsoft has not been defeated so much as it has walked off the battlefield.

Conclusion: The Ninth Generation Belongs to the PlayStation 5

The PlayStation 5 wins not through overwhelming brilliance, but through consistency in an era when its rivals began playing different games. It retained the clearest hardware identity, a unified performance target, strong third-party alignment, and enough significant releases to give the platform a recognizable shape.

Switch 2 may sell in vast numbers, but its success follows a logic different from the old performance race. Xbox Series remains relevant through services, though less clearly as the center of a hardware generation. By the older measure, the PS5 may therefore be the last obvious winner of the classic console contest.

In an industry drifting toward platform agnosticism, cloud integration, and hardware-service hybridization, the PS5 may well be the final console to win a generation in the old sense of the phrase.

Postscript: PS5’s Victory Is Real — but It Is Still a “Tall Dwarf Among Dwarfs”

Recognizing the PS5 as the generation’s winner does not require viewing it through rose-tinted glass. Its triumph is relative, not absolute.

The console itself carries notable flaws: a bulky chassis defined by thermal constraints, a cooling solution that grows noisy under prolonged load, and a GPU whose ray-tracing capability lags significantly behind contemporary PC standards. Its SSD expansion model, while flexible, lacks the elegant plug-and-play simplicity of cartridge-based ecosystems. And despite Sony’s immense studio power, the first-party pipeline slowed visibly mid-generation due to long development cycles and pandemic-driven delays.

That slowdown was compounded by an early-cycle misread inside Sony. Leadership placed heavy faith in live-service revenue, and even studios known for story-driven work were drawn into ongoing multiplayer projects outside their established strengths. Years of development were absorbed by service roadmaps just as budgets and schedules were already expanding. Concord’s collapse became the most visible sign of this detour: a reminder that directing too many studios toward “forever games” could also deprive PlayStation of the focused releases on which its identity had been built.

The PS5’s system software reflects similar unevenness—rigid UI navigation, limited customization, and a trophy ecosystem that has stagnated. Even PlayStation Plus, despite repeated restructuring, has yet to form a cohesive counterweight to Game Pass.

Sony won this generation because it made fewer strategic errors than its competitors, not because its machine embodies technical or conceptual perfection.
Its victory marks the epilogue of an era—an echo of a time when generations were clearly defined and console wars were fought on familiar terrain.

The PS5 did not stand especially tall; the ground around it sank. That is what makes it, perhaps, the last tall dwarf of the console age.

This article is a collected edition of the How the Box Decides series.

• From Dedicated Boxes to Swappable Software
• Curation, Characters, and Enhancement Chips
• When CDs Beat Cartridges
• DVD Players, Modems, and Hard Drives
• HD, Subscriptions, and Hybrids

:::

Introduction: Beyond “8-bit, 16-bit, 4K”

“Console generations” are usually drawn as a simple staircase of power:
8-bit → 16-bit → 3D → HD → 4K → ray tracing.

Every few months someone declares the war over — most recently, GameStop’s Oct 25, 2025 statement joked that the console wars started with Halo: Combat Evolved exclusivity and would end once Halo: Campaign Evolved lands on PlayStation in 2026 — yet the debates never quite die.

In reality, every step hides a much messier layer:

• media choices — cartridges vs CDs vs DVDs vs SSDs
• business friction — Nintendo vs Sony vs Philips vs SEGA vs Microsoft
• ecosystem bets — tight control vs open third parties vs subscription services

This article follows that history generation by generation, but more importantly, console by console. I linger on the individual machines because each of them altered the path in some particular way.

“Generation 0”: Nintendo’s Dedicated TV Boxes (1977–1980)

Color TV-Game Series — Nintendo Before Famicom

Before the Famicom, Nintendo’s first home consoles were a line of dedicated machines: the Color TV-Game series.
(Technically these shipped after early first-generation systems like the Magnavox Odyssey; I’m calling them out separately because they were Nintendo’s bridge from toys to fully fledged consoles.)

These were Japan-only boxes sold between 1977 and 1980, with no interchangeable games:

• Color TV-Game 6 and 15 (1977): Pong-style tennis variations with knobs on the console itself. Players turned dials to move paddles while the console drew simple colored blocks on the TV.
• Racing 112 (1978): A tiny “cockpit” console with a steering wheel and accelerator, rendering a top-down racing lane with moving obstacles.
• Block Kuzushi (1979): Nintendo’s take on Breakout, with a paddle controller and brick-breaking gameplay baked directly into the hardware.
• Computer TV Game (1980): A dedicated version of Nintendo’s own Computer Othello arcade game, turning your TV into a digital board game surface.

::: grid {cols=3,gap=4,type=images}

:::

These machines sold a few million units in Japan and did one crucial thing:
they taught Nintendo how to design TV-connected electronics, industrial enclosures, controllers, and game logic long before the Famicom ever existed.

They were sealed boxes, but they proved something bigger: there was a home for videogames in the living room.

First Generation: Magnavox Odyssey and the Concept (1972–1975)

Magnavox Odyssey — Videogames as Electronic Board Games

The Magnavox Odyssey (1972) is widely recognized as the first commercial home videogame console.

Technically, it barely resembles what came later:

• It used only analog circuitry — no CPU, no RAM.
• It output a few movable squares and a line; everything else was imagination.
• It shipped with plastic screen overlays, dice, cards, and paper money.

You didn’t “buy a game” as software; you got rule sheets telling you how to interpret those moving squares as tennis, hockey, or haunted houses.

Odyssey proved the core idea: a TV could be interactive.
Everything after it is refinement.

Second Generation: Cartridges and the Crash (1976–1983)

Fairchild Channel F — Swappable Software

The Fairchild Channel F was the first console with ROM cartridges you could plug in and out.

• Built around a Fairchild F8 CPU, it split games off from the hardware.
• Cartridges allowed developers to ship new experiences without changing the console itself.

Its library was modest and it lost the commercial race, but it introduced an architecture that every later console inherited:
a stable hardware base with evolving software on top.

Atari 2600 — The Rise and the Burnout

The Atari 2600 (1977) turned that architecture into a cultural phenomenon.

• Hardware: a tiny 8-bit CPU, 128 bytes of RAM, and a video chip that programmers had to “race” line-by-line to draw anything at all.
• Library: Space Invaders, Pitfall!, Adventure, early sports games — all crammed into cartridges as small as 2–4 KB.
• Business model: anyone could make a cartridge for it.

That openness created the first third-party wave (Activision comes from ex-Atari developers), but also a flood of low-quality games.
By 1983, the market was overloaded with bad cartridges; retailers dumped stock and pulled back shelf space. The North American videogame crash followed.

Atari 2600 proved two things at once:

1. A home console business could be huge.
2. Without control over software quality, it could collapse just as fast.

Third Generation: Famicom / NES Rebuilds the Market (1983–1989)

Nintendo Famicom / NES — Hardware + Licensing

Nintendo’s Family Computer (Famicom) (1983 Japan) and its western cousin NES (1985 North America) revived the console market on a very different model.

::: grid {cols=2,gap=4,type=images}

:::

• Hardware:
  • 8-bit CPU with a custom PPU (Picture Processing Unit) for sprites and smooth scrolling.
  • Simple but flexible audio with multiple channels for melodies, bass lines, and noise.
• Form factor:
  • In Japan, playful red-and-white plastic with hard-wired controllers.
  • In the US, redesigned as a VCR-like front-loader to distance it from “game console” stigma.

The key innovation wasn’t just technical. It was business:

• Nintendo created a strict licensing program.
• Every licensed cartridge used a lockout chip and carried the “Seal of Quality”.
• Third parties had quotas and contract limits; unlicensed publishers were pushed out.

This tight control avoided another Atari-style crash and let Nintendo curate a library that still defines the 8-bit era:

• Super Mario Bros. redesigned platforming physics.
• The Legend of Zelda introduced battery-backed save data on a console cartridge.
• Metroid and Castlevania experimented with nonlinear world design.

Sega Master System — Technically Strong, Ecologically Weak

The Sega Master System (based on the Mark III in Japan) often looked better on paper than the NES:

• It supported more on-screen colors.
• Its Z80 CPU was familiar and tried-and-true.
• Some ports (e.g. Space Harrier, OutRun) were closer to their arcade originals.

But Sega lacked Nintendo’s licensing network, and many third-party publishers were contractually tied to releasing only on Nintendo hardware in certain markets.
Master System thrived more in Europe and Brazil, but in North America it became a niche box — proof that hardware specs alone cannot win a generation.

Fourth Generation: 16-bit Wars and Early Optical Experiments (1988–1995)

Sega Mega Drive / Genesis — Speed, Attitude, and “What Nintendon’t”

The Mega Drive (Japan) / Genesis (North America) pushed Sega into the lead, at least briefly.

• Hardware:
  • Motorola 68000 CPU at 7.6 MHz for the main logic.
  • Z80 coprocessor and an FM synth audio chip (Yamaha YM2612) for arcade-like sound.
• Design goal:
  • Fast scrolling, big sprites, and responsive controls — ideal for action and sports.

Marketing in the US turned it into the “cool” console:

• Slogans like “Genesis does what Nintendon’t”.
• Edgier games (blood in Mortal Kombat with a code, versus censored SNES).
• Sonic the Hedgehog as a mascot built on sheer speed.

Genesis repositioned consoles from “toys” toward teen culture.

Super Nintendo (SNES) — Enhancement Chips and RPG Heaven

Nintendo answered with the Super Famicom / SNES.

• Hardware:
  • Slower main CPU than Genesis, but a powerful PPU capable of multiple layers and the famous Mode 7 (affine transform of background planes for pseudo-3D effects).
  • High-quality sample-based audio through the SPC700 chip.

Instead of betting everything on base hardware, Nintendo embraced enhancement chips on the cartridge:

• Super FX for polygonal 3D in games like Star Fox.
• DSP and other coprocessors for fast math, rotation, scaling, and special effects.

SNES became the home of lavish 16-bit JRPGs and action adventures:

• Final Fantasy IV/VI, Chrono Trigger, Secret of Mana, A Link to the Past.
• Each one pushing story, music, and system design into territory that still feels dense today.

NEC PC Engine / TurboGrafx-16 — Cards and CDs

NEC’s PC Engine (TurboGrafx-16 in the West) chose a different path:

• Games came on slim HuCards (credit-card-sized cartridges).
• An optional CD-ROM add-on brought CD audio and large storage to the living room earlier than most rivals.

It shined especially in Japan with shooters and action games, and showed how optical media could power animated cutscenes and CD music even on modest hardware.

Neo Geo AES — The Luxury Arcade in a Box

SNK’s Neo Geo AES was effectively an arcade board repackaged for home.

• It used the same architecture as the Neo Geo MVS arcade system.
• Cartridges were huge and expensive; individual games could cost more than entire competing consoles.

Owning a Neo Geo meant having nearly pixel-perfect versions of Metal Slug, King of Fighters, and Samurai Shodown at home — if you could afford it.
It demonstrated a different fantasy: no compromises, just pay whatever it costs.

Fifth Generation: 3D, Broken Deals, and the Cartridge/CD Split (1994–2000)

The Broken SNES CD-ROM Partnership

Late in the SNES era, Nintendo and Sony worked together on a SNES CD-ROM add-on, code-named “Play Station”.

• The contract reportedly gave Sony unusually strong control over the CD format and its licensing.
• Nintendo feared losing control of its software revenue.

In a now-famous move, Nintendo publicly announced at a trade show that it would instead partner with Philips on CD technology — right after Sony had presented the joint project.

Consequence:

• Sony continued development internally, turning the project into a standalone console: PlayStation.
• Philips received limited rights to use Nintendo characters, which eventually birthed the infamous low-budget Zelda and Mario games on CD-i.
• Nintendo cancelled the SNES CD path altogether.

Nintendo’s attempt to protect its licensing power accidentally created its next major competitor.

Sony PlayStation — Cheap CD Hardware, Friendly to Third Parties

The PlayStation launched in 1994 (Japan) and 1995 (US/EU).

::: grid {cols=2,gap=4,type=images}

:::

• Hardware:
  • 32-bit R3000 class CPU built for 3D math.
  • Affine-textured polygons with no hardware perspective correction — visually wobbly but very fast for the era.
  • Simple, effective audio for streamed CD music and sound effects.
• Media:
  • CD-ROM with ~650–700 MB capacity per disc, dirt-cheap to manufacture compared to cartridges.
• Business:
  • Licensing terms and dev tools were much friendlier than Nintendo’s.
  • Sony’s experience in music and movies gave it a distribution edge.

Third-party publishers flocked to PlayStation, especially those who wanted cinematic production values: pre-rendered backgrounds, FMV cutscenes, full voice tracks.

Nintendo 64 — Fast Cartridges, Slow Ecosystem

The Nintendo 64 answered with powerful 3D hardware:

• 64-bit CPU, a capable GPU, advanced anti-aliasing and texture filtering for its time.
• Four controller ports built in, encouraging local multiplayer.

But Nintendo chose to stick with ROM cartridges:

• Pros: instant loading, robust physical media, better control over piracy and distribution.
• Cons: high manufacturing cost per unit, much smaller storage than CDs, logistics risk for publishers.

As a result:

• Many third-party publishers were wary of tying up capital in expensive carts.
• FMV and voiced dialogue were constrained by capacity.
• Big cinematic projects gravitated toward PlayStation instead.

N64 still produced landmark design work — Super Mario 64, Ocarina of Time, GoldenEye 007 — but commercially, it ceded the mass-market lead to Sony.

iQue Player — A Late N64 for Mainland China

With traditional consoles effectively banned in mainland China during the late ’90s and early ’00s, Nintendo partnered with local firm iQue (founded by Dr. Wei Yen) to release the iQue Player in 2003.
The device squeezed N64 silicon into a controller-sized shell that plugged directly into a TV and drew games from flash memory.

• Titles were purchased via kiosk downloads in Chinese shopping malls, avoiding boxed media and helping Nintendo satisfy local regulations.
• The library leaned on N64 classics such as Super Mario 64, Ocarina of Time, and Star Fox 64, all localized into simplified Chinese for the first time.
• Save data lived inside the controller and could be backed up to memory cards, a nod to the system’s kiosk-driven software flow.

iQue Player arrived years after the N64’s global run, but it highlighted how hardware makers contorted their designs to reach huge but restricted markets — a reminder that “generation” boundaries look very different outside North America and Japan.

Why Final Fantasy VII Left Nintendo

Square had a long history with Nintendo. But when it came to Final Fantasy VII, they jumped.

• Nintendo 64 cartridges topped out around 64 MB for commercial releases.
• A single FFVII disc on PS1 held around 700 MB, and the game shipped on three discs.
• Most of that space went to CG cinematics and audio — exactly the things Square wanted to lean on for a more cinematic FF.

Square’s leadership has since been explicit: the choice of PlayStation was driven by CD-ROM capacity and flexibility, not just raw polygon power.
On N64, FFVII would have required heavy cuts to the cinematic content; on PS1, it could be the sprawling CG-driven RPG they envisioned.

This one decision symbolized a broader migration:
RPGs and many narrative-heavy games followed the storage.

Sega Saturn — 2D Power, 3D Confusion

The Saturn was Sega’s 32-bit machine:

• Architected around dual CPUs and multiple coprocessors.
• Exceptionally strong at 2D sprite handling and certain 3D workloads.

But:

• The multi-CPU design was hard to optimize for; dev tools lagged.
• Sega’s prior missteps with add-ons (Mega-CD, 32X) had already shaken retailer confidence.
• The surprise early launch in North America angered some retail partners.

Saturn found a loyal niche (especially in Japan), but could not match PlayStation’s momentum or library breadth.

Victor V-Saturn — Licensed Twin with Cosmetic Flair

To widen retail reach inside Japan, Sega licensed the Saturn hardware to major electronics partners. Victor (JVC) shipped the V-Saturn line in 1995 under model numbers RG-JX1 and later RG-JX2.

• Internally, these machines mirrored Sega’s revisions; the differences were aesthetic and branding focused.
• A darker chassis, Victor logos, and bespoke boot animations gave the console a hi-fi vibe, and some bundles included Victor-branded controllers.
• Because the BIOS tweaks were cosmetic, V-Saturn supported the full Saturn software library and peripherals with no compatibility caveats.

Collectors chase the V-Saturn today not for new capabilities, but because it captures the brief moment when Sega tried letting trusted manufacturing partners sell “clone” hardware to keep momentum in its home market.

Philips CD-i — Licensed Characters, No Real Direction

The CD-i wasn’t a traditional console; Philips marketed it as a “multimedia player”.

:::grid {cols=2,gap=4,type=images}

:::

• Its hardware targeted video and simple interactivity for educational titles.
• As part of its deal with Nintendo, Philips published a handful of officially licensed Zelda and Mario games.

Those games were notorious — clunky controls, awkward animation, minimal oversight from Nintendo.
CD-i showed that IP licensing without platform focus doesn’t create a successful console.

Sixth Generation: PS2 Peak, Dreamcast Exit, Xbox Arrives (1998–2013)

Sega Dreamcast — First Out, First to Fall

The Dreamcast launched in 1998 (JP) / 1999 (NA/EU).

• Hardware:
  • Hitachi SH-4 CPU and a PowerVR2 GPU, making it very efficient at 3D rendering.
  • Visual quality close to contemporary arcade boards.
• Media & Online:
  • Proprietary GD-ROM discs (~1 GB).
  • Built-in modem for dial-up online play; Phantasy Star Online became one of the first console online RPGs.
  • Early experiments with downloadable content.

Dreamcast felt like a machine from the future, but Sega was weakened financially, and the looming PS2 hype (plus rampant piracy later) squeezed it out.

In 2001, Sega discontinued Dreamcast and exited the console hardware business, pivoting to third-party publishing.

Sony PlayStation 2 — DVD Player of the World

The PS2 (2000) is often called the most successful console ever for a reason:

:::grid {cols=2,gap=4,type=images}

:::

• It played PS2 games, PS1 games, and DVD movies out of the box.
• For many households, it was the first affordable DVD player.
• It built on the existing PlayStation developer community and third-party relationships.

Technically, the “Emotion Engine” CPU and GS GPU were quirky but powerful in the right hands.
Commercially, the combination of backward compatibility, huge library, and media playback made PS2 the obvious default choice for an entire generation.

PSX — DVR Ambition Built on PS2 Silicon

In 2003 Sony tried to fuse home video recording with the PlayStation brand via the Japan-only PSX (models DESR-5000/7000 and later revisions).

• Under the hood it was effectively a PlayStation 2 motherboard paired with large hard drives, analog TV tuners, and DVD burning for archiving shows.
• The machine debuted the XrossMediaBar interface that later appeared on PSP and PS3, along with features like Memory Stick slots and in-home video editing.
• High launch prices (¥79,800 and up), sluggish firmware updates, and the rise of cheaper DVRs kept sales modest, and Sony never exported it beyond Japan.

PSX foreshadowed the “all-in-one living-room hub” pitch that would resurface in later generations, even if the hardware itself became a niche collector piece.

Nintendo GameCube — Compact, Efficient, and Overshadowed

The GameCube (2001) was Nintendo’s small cube with a comfort-focused controller.

• Hardware:
  • IBM “Gekko” CPU and ATI “Flipper” GPU — straightforward, potent hardware for the time.
  • Mini-DVD discs (1.5 GB) that balanced capacity, load times, and piracy resistance.
• Strengths:
  • Very strong first-party titles: Metroid Prime, Super Smash Bros. Melee, Wind Waker, F-Zero GX.
  • Excellent analog triggers and ergonomics on the controller.

Yet in the marketplace, it struggled against PS2’s overwhelming momentum and the allure of DVD playback.
GameCube felt like “the enthusiast’s second console” more than the primary system.

Panasonic Q — GameCube Meets Living-Room AV

Nintendo licensed the GameCube hardware to Panasonic (then still Matsushita) for the Japan-only Panasonic Q (2001).

• It packed standard GameCube guts inside a brushed-metal chassis with mirrored front panel, adding DVD-Video playback and beefier home-theater audio outputs.
• A backlit front display, bundled remote, and multi-region capabilities positioned it as a premium AV component rather than a toy-like cube.
• Pricing landed around ¥39,800—more than double a regular GameCube—which, combined with its Japan-only release, kept sales niche.

Panasonic Q showed Nintendo experimenting with partnerships to answer the DVD feature gap without redesigning the core console, but the cost premium proved that style and functionality alone can’t overcome market momentum.

Microsoft Xbox — A PC in Console Clothing

The original Xbox (2001) was a direct incursion from Microsoft:

• x86 CPU, NVIDIA GPU, and an internal hard drive.
• Built to look and feel like a console but think like a PC.

Its most important contribution was Xbox Live:

• Unified online service, friends list, voice chat, digital content.
• Halo turned into the definitive console FPS experience.

Xbox didn’t win the generation in raw sales, but it defined a networked services model that everyone else would eventually follow.

Seventh Generation: HD, Motion Controls, and Early “Games as Services” (2005–2017)

Xbox 360 — Fast Start, Red Ring

The Xbox 360 (2005) hit first in the HD race.

:::grid {cols=3,gap=4,type=images}

:::

• Strong GPU and flexible architecture made it easy to target.
• Standardized achievements and a robust Xbox Live experience shaped the modern notion of a console account ecosystem.
• Many cross-platform games ran best or first on 360 in the early years.

Its weakness was hardware reliability: early models notoriously suffered from the “Red Ring of Death”.
Microsoft’s extended warranty and repair program was costly, but preserved user trust enough to keep 360 competitive for the generation’s entire span.

PlayStation 3 — Complex Silicon, Slow Burn

The PS3 (2006) aimed high:

:::grid {cols=3,gap=4,type=images}

:::

• The Cell processor was exotic and powerful, but notoriously hard to exploit fully.
• Blu-ray drive expanded disc capacity and helped win the HD optical media war.

At launch, PS3 was:

• Expensive.
• Short on must-have exclusives.
• Burdened with a developer-unfriendly reputation.

Over time, revised hardware, lower prices, and strong first-party titles — Uncharted, Killzone 2, The Last of Us — turned it around.

A curious side note:
due to a patent dispute with Immersion, the PS3 launched with SIXAXIS controllers that lacked rumble, at the exact moment when HD graphics made rumble more impactful.
Only after settling did Sony ship the DualShock 3, restoring vibration alongside motion sensing.

Nintendo Wii — Opting Out of the Power Race

The Wii (2006) used hardware derived from GameCube, significantly weaker than 360 and PS3.
It didn’t matter.

• The Wii Remote’s motion controls invited non-gamers into the experience.
• Wii Sports and Wii Fit turned the console into a living-room fitness and party machine.
• It became a phenomenon at family gatherings, rehab centers, and retirement homes.

Wii sold over 100 million units, but its modest power and SD output meant it could not sustain the same third-party HD pipeline as 360/PS3.
Still, it proved decisively that “different” can beat “more powerful”.

Eighth Generation: HD Course Corrections Before the SSD Era (2012–2017)

Wii U — The Misunderstood Bridge

The Wii U (2012) tried to build on the Wii’s user base with a GamePad featuring a built-in screen.

In theory:

• It could act as an asymmetric second screen in games.
• It enabled off-TV play in some titles.

In practice:

• The branding was confusing — many consumers thought it was just an add-on for Wii.
• Third-party support was weak, and the hardware was a half-step below PS4/Xbox One.

Wii U underperformed badly, but its experiments with hybrid play laid conceptual groundwork for the Switch.

PlayStation 4 — Back to Basics, and It Worked

The PS4 (2013) was Sony’s course correction.

:::grid {cols=3,gap=4,type=images}

:::

• Architecturally, it was a PC-like x86-64 APU with a straightforward GPU and fast unified GDDR5 memory.
• Marketing messaged a focus on “for the players”, not as an all-purpose media device.

Developers found it easy to work with, and players got:

• A strong lineup of exclusives.
• Solid third-party performance.
• Reasonable pricing.

PS4 became the default console for many players worldwide and a central pillar for AAA and indie alike.

Xbox One — TV First, Games Second (Then Back Again)

:::grid {cols=3,gap=4,type=images}

:::

The Xbox One (2013) initially chased an all-in-one living-room vision:

• Heavy integration with live TV.
• Mandatory Kinect bundling.
• Announced restrictions on used games and online authentication.

The response was negative enough that Microsoft rolled back key policies before launch and eventually dropped Kinect as mandatory.

Over the generation, Xbox pivoted to:

• Aggressive backward compatibility for older Xbox titles.
• Xbox Game Pass, a subscription service offering a rotating catalog of games for a flat fee.

By the end of the generation, Xbox’s identity had shifted from individual console sales to service-centric engagement.

This generation set the table for the solid-state era: Sony and Microsoft concentrated on HD performance and developer goodwill, while Nintendo’s half-step experiments exposed the appetite for hybrid play that would define the next wave.

Ninth Generation: Hybrid Becomes Mainstream + SSD Leap (2017–2025)

Analysts and tech outlets such as Tom’s Guide and Pocket-lint (as summarized on Wikipedia’s ninth-generation overview) have noted that the Nintendo Switch competes directly with PS5 and Xbox Series despite arriving earlier and using weaker silicon. Following that research, this article treats 2017’s Switch launch as the opening shot of the ninth generation, with Sony and Microsoft’s 2020 hardware doubling down on SSD-first design philosophy.

Nintendo Switch — Hybrid Done Right

The Nintendo Switch (2017) solved the “what is this thing?” problem that Wii U had failed to address.

:::grid {cols=3,gap=4,type=images}

:::

• A tablet-like main unit with detachable Joy-Con controllers.
• Dock mode for TV play, handheld mode for portable play, tabletop mode for quick multiplayer.
• Modest Tegra-based hardware that prioritized efficiency over raw power.

With The Legend of Zelda: Breath of the Wild as a launch title and Super Mario Odyssey following shortly, Switch proved you could:

• Have big, open-world experiences both on TV and on the go.
• Accept weaker specs in exchange for flexibility and first-party strength.

Switch blurred the lines between home console and handheld in a way that stuck, and its 2021 OLED refresh kept it in conversation with PS5 and Series consoles throughout their shortage-plagued first years.

PlayStation 5 — I/O as a Design Constraint

The PS5 (2020) reoriented console design around storage bandwidth as much as compute:

• A custom NVMe SSD and demarcated I/O pipeline drastically cut load times.
• Developers could treat streaming as a constant flow rather than a stuttering bottleneck.
• Hardware ray tracing and up to 4K/120 Hz output supported high-end visual targets.

It also introduced the DualSense controller:

• Adaptive triggers that change resistance dynamically (bow tension, gun triggers).
• High-resolution haptics that can simulate texture and rhythm more precisely than older rumble motors.

PS5 stayed with the PS4’s general philosophy — powerful but developer-friendly hardware, strong exclusives, and large third-party support — but removed “waiting” as a core part of the experience.

Xbox Series X|S — Hardware Pair + Service First

The Xbox Series X and Series S (2020) doubled down on Microsoft’s ecosystem strategy.

• Series X: high-end 4K-focused box.
• Series S: cheaper, digital-only 1440p-class machine.

Common threads:

• Extensive backward compatibility, often with auto-enhancements.
• Tight integration with Game Pass, positioning the console as the easiest way to access a large game library rather than as a standalone silo.

In this model, your Xbox account and subscription matter more than the specific console model you own.

Tenth Generation: Switch 2 Opens the Post-Hybrid Decade (2025– )

Nintendo Switch 2 — Fastest Start in Console History

The Nintendo Switch 2 launched on June 5, 2025, as the successor to the original Switch.

Core ideas:

• It keeps the hybrid form factor: handheld unit plus dock for TV output.
• Hardware is significantly more capable, allowing higher resolutions and smoother performance while staying portable.
• Nintendo positions it as a continuation, not a reboot — your understanding of “a Switch” still applies.

Commercially, its opening was explosive:

• Over 3.5 million units sold worldwide in the first four days.
• The fastest-selling Nintendo hardware launch ever.
• Analysts and multiple outlets have described it as the fastest-selling console launch in industry history, beating even PS4 and PS5’s comparable early milestones.

Switch 2 didn’t join a traditional power race with PS5 and Series X/S.
Instead, it extended the same trade-off Nintendo has favored since Wii:

• Accept lower peak specs.
• Push on form factor, accessibility, and first-party software.
• Let the other two fight over teraflops — just now from a clean, next-generation slate.

Closing: What Generations Were Really About

If you connect all these machines, you don’t just get a power curve. You get a chain of decisions:

• Atari 2600 showed what happens when you open a platform fully and never say “no”.
• Famicom/NES showed that tight licensing and curation could rebuild a broken market.
• Nintendo’s attempt to preserve control over SNES CD deals essentially created PlayStation as a rival.
• Sticking with cartridges on N64 protected margins but cost them Square and a generation of cinematic RPGs.
• Sega’s repeated hardware pivots, culminating in Saturn and Dreamcast, exhausted both finances and trust.
• Microsoft reframed consoles as networked services, culminating in Game Pass and deep backward compatibility.
• Sony rode optical media, then HD, then SSD-driven design, staying focused on developer support and single-player epics.
• Nintendo stepped sideways: from dedicated TV boxes to Famicom, from Wii’s motion controls to the hybrid Switch and Switch 2.

You can read console generations as an arms race in CPU and GPU numbers.
Or you can read them as a long series of answers to a quieter question:

What do we want “playing at home” to feel like —
a locked box, a media center, a subscription, a portable window, or something in between?

The hardware keeps changing.
The arguments — about media formats, control, ecosystems, and who gets paid — never really stopped.
Every time you pick up a controller today, you’re feeling the outcome of fifty years of those arguments, condensed into a single plastic shell and a startup chime.

Further Reading

• Consollection.de — A multilingual catalog that documents console hardware variants, prototypes, and accessories.
• The Strong Museum timeline — A succinct timeline of landmark home systems across five decades.
• RetroRGB — Hardware repair, video-output mods, and preservation tips for dozens of platforms.
• Console5 Wiki — Schematics, capacitor lists, and repair guides for classic consoles from Atari through Sega and beyond.

• accounts and storefronts
• patches and DLC
• online identity and social layers
• long-tail monetization and subscriptions

At the same time, Nintendo proved (again and again) that the winning move isn’t always to fight the same war.

HD Becomes the Default (and Reliability Matters)

Xbox 360 — Fast Start, Red Ring

The Xbox 360 (2005) hit first in the HD race.

:::grid {cols=3,gap=4,type=images}

:::

• Strong GPU and flexible architecture made it easy to target.
• Standardized achievements and a robust Xbox Live experience shaped the modern notion of a console account ecosystem.
• Many cross-platform games ran best or first on 360 in the early years.

Its weakness was hardware reliability: early models notoriously suffered from the “Red Ring of Death”.
Microsoft’s extended warranty and repair program was costly, but preserved user trust enough to keep 360 competitive for the generation’s entire span.

PlayStation 3 — Complex Silicon, Slow Burn

The PS3 (2006) aimed high:

:::grid {cols=3,gap=4,type=images}

:::

• The Cell processor was exotic and powerful, but notoriously hard to exploit fully.
• Blu-ray drive expanded disc capacity and helped win the HD optical media war.

At launch, PS3 was:

• Expensive.
• Short on must-have exclusives.
• Burdened with a developer-unfriendly reputation.

Over time, revised hardware, lower prices, and strong first-party titles — Uncharted, Killzone 2, The Last of Us — turned it around.

A curious side note:
due to a patent dispute with Immersion, the PS3 launched with SIXAXIS controllers that lacked rumble, at the exact moment when HD graphics made rumble more impactful.
Only after settling did Sony ship the DualShock 3, restoring vibration alongside motion sensing.

When “Different” Beats “More Powerful”

Nintendo Wii — Opting Out of the Power Race

The Wii (2006) used hardware derived from GameCube, significantly weaker than 360 and PS3.
It didn’t matter.

• The Wii Remote’s motion controls invited non-gamers into the experience.
• Wii Sports and Wii Fit turned the console into a living-room fitness and party machine.
• It became a phenomenon at family gatherings, rehab centers, and retirement homes.

Wii sold over 100 million units, but its modest power and SD output meant it could not sustain the same third-party HD pipeline as 360/PS3.
Still, it proved decisively that “different” can beat “more powerful”.

Course Corrections, Digital Stores, and Subscriptions

Wii U — The Misunderstood Bridge

The Wii U (2012) tried to build on the Wii’s user base with a GamePad featuring a built-in screen.

In theory:

• It could act as an asymmetric second screen in games.
• It enabled off-TV play in some titles.

In practice:

• The branding was confusing — many consumers thought it was just an add-on for Wii.
• Third-party support was weak, and the hardware was a half-step below PS4/Xbox One.

Wii U underperformed badly, but its experiments with hybrid play laid conceptual groundwork for the Switch.

PlayStation 4 — Back to Basics, and It Worked

The PS4 (2013) was Sony’s course correction.

:::grid {cols=3,gap=4,type=images}

:::

• Architecturally, it was a PC-like x86-64 APU with a straightforward GPU and fast unified GDDR5 memory.
• Marketing messaged a focus on “for the players”, not as an all-purpose media device.

Developers found it easy to work with, and players got:

• A strong lineup of exclusives.
• Solid third-party performance.
• Reasonable pricing.

PS4 became the default console for many players worldwide and a central pillar for AAA and indie alike.

Xbox One — TV First, Games Second (Then Back Again)

:::grid {cols=3,gap=4,type=images}

:::

The Xbox One (2013) initially chased an all-in-one living-room vision:

• Heavy integration with live TV.
• Mandatory Kinect bundling.
• Announced restrictions on used games and online authentication.

The response was negative enough that Microsoft rolled back key policies before launch and eventually dropped Kinect as mandatory.

Over the generation, Xbox pivoted to:

• Aggressive backward compatibility for older Xbox titles.
• Xbox Game Pass, a subscription service offering a rotating catalog of games for a flat fee.

By the end of the generation, Xbox’s identity had shifted from individual console sales to service-centric engagement.

This generation set the table for the solid-state era: Sony and Microsoft concentrated on HD performance and developer goodwill, while Nintendo’s half-step experiments exposed the appetite for hybrid play that would define the next wave.

Hybrid and SSD: Two Paths Forward (2017–2025)

Analysts and tech outlets such as Tom’s Guide and Pocket-lint (as summarized on Wikipedia’s ninth-generation overview) have noted that the Nintendo Switch competes directly with PS5 and Xbox Series despite arriving earlier and using weaker silicon. Following that research, this article treats 2017’s Switch launch as the opening shot of the ninth generation, with Sony and Microsoft’s 2020 hardware doubling down on SSD-first design philosophy.

Nintendo Switch — Hybrid Done Right

The Nintendo Switch (2017) solved the “what is this thing?” problem that Wii U had failed to address.

:::grid {cols=3,gap=4,type=images}

:::

• A tablet-like main unit with detachable Joy-Con controllers.
• Dock mode for TV play, handheld mode for portable play, tabletop mode for quick multiplayer.
• Modest Tegra-based hardware that prioritized efficiency over raw power.

With The Legend of Zelda: Breath of the Wild as a launch title and Super Mario Odyssey following shortly, Switch proved you could:

• Have big, open-world experiences both on TV and on the go.
• Accept weaker specs in exchange for flexibility and first-party strength.

Switch blurred the lines between home console and handheld in a way that stuck, and its 2021 OLED refresh kept it in conversation with PS5 and Series consoles throughout their shortage-plagued first years.

PlayStation 5 — I/O as a Design Constraint

The PS5 (2020) reoriented console design around storage bandwidth as much as compute:

• A custom NVMe SSD and demarcated I/O pipeline drastically cut load times.
• Developers could treat streaming as a constant flow rather than a stuttering bottleneck.
• Hardware ray tracing and up to 4K/120 Hz output supported high-end visual targets.

It also introduced the DualSense controller:

• Adaptive triggers that change resistance dynamically (bow tension, gun triggers).
• High-resolution haptics that can simulate texture and rhythm more precisely than older rumble motors.

PS5 stayed with the PS4’s general philosophy — powerful but developer-friendly hardware, strong exclusives, and large third-party support — but removed “waiting” as a core part of the experience.

Xbox Series X|S — Hardware Pair + Service First

The Xbox Series X and Series S (2020) doubled down on Microsoft’s ecosystem strategy.

• Series X: high-end 4K-focused box.
• Series S: cheaper, digital-only 1440p-class machine.

Common threads:

• Extensive backward compatibility, often with auto-enhancements.
• Tight integration with Game Pass, positioning the console as the easiest way to access a large game library rather than as a standalone silo.

In this model, your Xbox account and subscription matter more than the specific console model you own.

Switch 2 and the Post-Hybrid Decade (2025– )

Nintendo Switch 2 — Fastest Start in Console History

The Nintendo Switch 2 launched on June 5, 2025, as the successor to the original Switch.

Core ideas:

• It keeps the hybrid form factor: handheld unit plus dock for TV output.
• Hardware is significantly more capable, allowing higher resolutions and smoother performance while staying portable.
• Nintendo positions it as a continuation, not a reboot — your understanding of “a Switch” still applies.

Commercially, its opening was explosive:

• Over 3.5 million units sold worldwide in the first four days.
• The fastest-selling Nintendo hardware launch ever.
• Analysts and multiple outlets have described it as the fastest-selling console launch in industry history, beating even PS4 and PS5’s comparable early milestones.

Switch 2 didn’t join a traditional power race with PS5 and Series X/S.
Instead, it extended the same trade-off Nintendo has favored since Wii:

• Accept lower peak specs.
• Push on form factor, accessibility, and first-party software.
• Let the other two fight over teraflops — just now from a clean, next-generation slate.

Closing: What “Generations” Were Really About

If you connect all these machines, you don’t just get a power curve. You get a chain of decisions:

• Atari 2600 showed what happens when you open a platform fully and never say “no”.
• Famicom/NES showed that tight licensing and curation could rebuild a broken market.
• Nintendo’s attempt to preserve control over SNES CD deals essentially created PlayStation as a rival.
• Sticking with cartridges on N64 protected margins but cost them Square and a generation of cinematic RPGs.
• Sega’s repeated hardware pivots, culminating in Saturn and Dreamcast, exhausted both finances and trust.
• Microsoft reframed consoles as networked services, culminating in Game Pass and deep backward compatibility.
• Sony rode optical media, then HD, then SSD-driven design, staying focused on developer support and single-player epics.
• Nintendo stepped sideways: from dedicated TV boxes to Famicom, from Wii’s motion controls to the hybrid Switch and Switch 2.

You can read console generations as an arms race in CPU and GPU numbers.
Or you can read them as a long series of answers to a quieter question:

What do we want “playing at home” to feel like —
a locked box, a media center, a subscription, a portable window, or something in between?

The hardware keeps changing.
The arguments — about media formats, control, ecosystems, and who gets paid — never really stopped.
Every time you pick up a controller today, you’re feeling the outcome of fifty years of those arguments, condensed into a single plastic shell and a startup chime.

Further Reading

• Consollection.de — A multilingual catalog that documents console hardware variants, prototypes, and accessories.
• The Strong Museum timeline — A succinct timeline of landmark home systems across five decades.
• RetroRGB — Hardware repair, video-output mods, and preservation tips for dozens of platforms.
• Console5 Wiki — Schematics, capacitor lists, and repair guides for classic consoles from Atari through Sega and beyond.

• an appliance (DVD player, DVR, AV component)
• a network terminal (accounts, friends lists, voice chat, downloadable content)

This is where the modern expectation of “a console has an OS and a service layer” really begins.

Online Arrives Early, Then Leaves

Sega Dreamcast — First Out, First to Fall

The Dreamcast launched in 1998 (JP) / 1999 (NA/EU).

• Hardware:
  • Hitachi SH-4 CPU and a PowerVR2 GPU, making it very efficient at 3D rendering.
  • Visual quality close to contemporary arcade boards.
• Media & Online:
  • Proprietary GD-ROM discs (~1 GB).
  • Built-in modem for dial-up online play; Phantasy Star Online became one of the first console online RPGs.
  • Early experiments with downloadable content.

Dreamcast felt like a machine from the future, but Sega was weakened financially, and the looming PS2 hype (plus rampant piracy later) squeezed it out.

In 2001, Sega discontinued Dreamcast and exited the console hardware business, pivoting to third-party publishing.

The DVD Trojan Horse

Sony PlayStation 2 — DVD Player of the World

The PS2 (2000) is often called the most successful console ever for a reason:

:::grid {cols=2,gap=4,type=images}

:::

• It played PS2 games, PS1 games, and DVD movies out of the box.
• For many households, it was the first affordable DVD player.
• It built on the existing PlayStation developer community and third-party relationships.

Technically, the “Emotion Engine” CPU and GS GPU were quirky but powerful in the right hands.
Commercially, the combination of backward compatibility, huge library, and media playback made PS2 the obvious default choice for an entire generation.

PSX — DVR Ambition Built on PS2 Silicon

In 2003 Sony tried to fuse home video recording with the PlayStation brand via the Japan-only PSX (models DESR-5000/7000 and later revisions).

• Under the hood it was effectively a PlayStation 2 motherboard paired with large hard drives, analog TV tuners, and DVD burning for archiving shows.
• The machine debuted the XrossMediaBar interface that later appeared on PSP and PS3, along with features like Memory Stick slots and in-home video editing.
• High launch prices (¥79,800 and up), sluggish firmware updates, and the rise of cheaper DVRs kept sales modest, and Sony never exported it beyond Japan.

PSX foreshadowed the “all-in-one living-room hub” pitch that would resurface in later generations, even if the hardware itself became a niche collector piece.

Nintendo’s Hardware: Focused, But Not the Default

Nintendo GameCube — Compact, Efficient, and Overshadowed

The GameCube (2001) was Nintendo’s small cube with a comfort-focused controller.

• Hardware:
  • IBM “Gekko” CPU and ATI “Flipper” GPU — straightforward, potent hardware for the time.
  • Mini-DVD discs (1.5 GB) that balanced capacity, load times, and piracy resistance.
• Strengths:
  • Very strong first-party titles: Metroid Prime, Super Smash Bros. Melee, Wind Waker, F-Zero GX.
  • Excellent analog triggers and ergonomics on the controller.

Yet in the marketplace, it struggled against PS2’s overwhelming momentum and the allure of DVD playback.
GameCube felt like “the enthusiast’s second console” more than the primary system.

Panasonic Q — GameCube Meets Living-Room AV

Nintendo licensed the GameCube hardware to Panasonic (then still Matsushita) for the Japan-only Panasonic Q (2001).

• It packed standard GameCube guts inside a brushed-metal chassis with mirrored front panel, adding DVD-Video playback and beefier home-theater audio outputs.
• A backlit front display, bundled remote, and multi-region capabilities positioned it as a premium AV component rather than a toy-like cube.
• Pricing landed around ¥39,800—more than double a regular GameCube—which, combined with its Japan-only release, kept sales niche.

Panasonic Q showed Nintendo experimenting with partnerships to answer the DVD feature gap without redesigning the core console, but the cost premium proved that style and functionality alone can’t overcome market momentum.

The Service Layer Becomes Real

Microsoft Xbox — A PC in Console Clothing

The original Xbox (2001) was a direct incursion from Microsoft:

• x86 CPU, NVIDIA GPU, and an internal hard drive.
• Built to look and feel like a console but think like a PC.

Its most important contribution was Xbox Live:

• Unified online service, friends list, voice chat, digital content.
• Halo turned into the definitive console FPS experience.

Xbox didn’t win the generation in raw sales, but it defined a networked services model that everyone else would eventually follow.

What Comes Next

By the mid-2000s, the next baseline shift was obvious: HD TVs were everywhere, and the service layer was no longer optional.

Next: HD, Subscriptions, and Hybrids

• CDs were cheap to manufacture.
• CDs held orders of magnitude more data than cartridges.
• CDs lowered the financial risk for publishers — and enabled bigger, more cinematic games.

Once that equation landed, it didn’t just change graphics. It rearranged the entire industry’s alliances.

The Broken Deal: Nintendo, Sony, Philips

The Broken SNES CD-ROM Partnership

Late in the SNES era, Nintendo and Sony worked together on a SNES CD-ROM add-on, code-named “Play Station”.

• The contract reportedly gave Sony unusually strong control over the CD format and its licensing.
• Nintendo feared losing control of its software revenue.

In a now-famous move, Nintendo publicly announced at a trade show that it would instead partner with Philips on CD technology — right after Sony had presented the joint project.

Consequence:

• Sony continued development internally, turning the project into a standalone console: PlayStation.
• Philips received limited rights to use Nintendo characters, which eventually birthed the infamous low-budget Zelda and Mario games on CD-i.
• Nintendo cancelled the SNES CD path altogether.

Nintendo’s attempt to protect its licensing power accidentally created its next major competitor.

CD Economics + 3D Momentum: PlayStation

Sony PlayStation — Cheap CD Hardware, Friendly to Third Parties

The PlayStation launched in 1994 (Japan) and 1995 (US/EU).

::: grid {cols=2,gap=4,type=images}

:::

• Hardware:
  • 32-bit R3000 class CPU built for 3D math.
  • Affine-textured polygons with no hardware perspective correction — visually wobbly but very fast for the era.
  • Simple, effective audio for streamed CD music and sound effects.
• Media:
  • CD-ROM with ~650–700 MB capacity per disc, dirt-cheap to manufacture compared to cartridges.
• Business:
  • Licensing terms and dev tools were much friendlier than Nintendo’s.
  • Sony’s experience in music and movies gave it a distribution edge.

Third-party publishers flocked to PlayStation, especially those who wanted cinematic production values: pre-rendered backgrounds, FMV cutscenes, full voice tracks.

Cartridges Hold On — And Pay the Price

Nintendo 64 — Fast Cartridges, Slow Ecosystem

The Nintendo 64 answered with powerful 3D hardware:

• 64-bit CPU, a capable GPU, advanced anti-aliasing and texture filtering for its time.
• Four controller ports built in, encouraging local multiplayer.

But Nintendo chose to stick with ROM cartridges:

• Pros: instant loading, robust physical media, better control over piracy and distribution.
• Cons: high manufacturing cost per unit, much smaller storage than CDs, logistics risk for publishers.

As a result:

• Many third-party publishers were wary of tying up capital in expensive carts.
• FMV and voiced dialogue were constrained by capacity.
• Big cinematic projects gravitated toward PlayStation instead.

N64 still produced landmark design work — Super Mario 64, Ocarina of Time, GoldenEye 007 — but commercially, it ceded the mass-market lead to Sony.

iQue Player — A Late N64 for Mainland China

With traditional consoles effectively banned in mainland China during the late ’90s and early ’00s, Nintendo partnered with local firm iQue (founded by Dr. Wei Yen) to release the iQue Player in 2003.
The device squeezed N64 silicon into a controller-sized shell that plugged directly into a TV and drew games from flash memory.

• Titles were purchased via kiosk downloads in Chinese shopping malls, avoiding boxed media and helping Nintendo satisfy local regulations.
• The library leaned on N64 classics such as Super Mario 64, Ocarina of Time, and Star Fox 64, all localized into simplified Chinese for the first time.
• Save data lived inside the controller and could be backed up to memory cards, a nod to the system’s kiosk-driven software flow.

iQue Player arrived years after the N64’s global run, but it highlighted how hardware makers contorted their designs to reach huge but restricted markets — a reminder that “generation” boundaries look very different outside North America and Japan.

Why Final Fantasy VII Left Nintendo

Square had a long history with Nintendo. But when it came to Final Fantasy VII, they jumped.

• Nintendo 64 cartridges topped out around 64 MB for commercial releases.
• A single FFVII disc on PS1 held around 700 MB, and the game shipped on three discs.
• Most of that space went to CG cinematics and audio — exactly the things Square wanted to lean on for a more cinematic FF.

Square’s leadership has since been explicit: the choice of PlayStation was driven by CD-ROM capacity and flexibility, not just raw polygon power.
On N64, FFVII would have required heavy cuts to the cinematic content; on PS1, it could be the sprawling CG-driven RPG they envisioned.

This one decision symbolized a broader migration:
RPGs and many narrative-heavy games followed the storage.

Complexity, Confidence, and Misfires: Saturn and CD-i

Sega Saturn — 2D Power, 3D Confusion

The Saturn was Sega’s 32-bit machine:

• Architected around dual CPUs and multiple coprocessors.
• Exceptionally strong at 2D sprite handling and certain 3D workloads.

But:

• The multi-CPU design was hard to optimize for; dev tools lagged.
• Sega’s prior missteps with add-ons (Mega-CD, 32X) had already shaken retailer confidence.
• The surprise early launch in North America angered some retail partners.

Saturn found a loyal niche (especially in Japan), but could not match PlayStation’s momentum or library breadth.

Victor V-Saturn — Licensed Twin with Cosmetic Flair

To widen retail reach inside Japan, Sega licensed the Saturn hardware to major electronics partners. Victor (JVC) shipped the V-Saturn line in 1995 under model numbers RG-JX1 and later RG-JX2.

• Internally, these machines mirrored Sega’s revisions; the differences were aesthetic and branding focused.
• A darker chassis, Victor logos, and bespoke boot animations gave the console a hi-fi vibe, and some bundles included Victor-branded controllers.
• Because the BIOS tweaks were cosmetic, V-Saturn supported the full Saturn software library and peripherals with no compatibility caveats.

Collectors chase the V-Saturn today not for new capabilities, but because it captures the brief moment when Sega tried letting trusted manufacturing partners sell “clone” hardware to keep momentum in its home market.

Philips CD-i — Licensed Characters, No Real Direction

The CD-i wasn’t a traditional console; Philips marketed it as a “multimedia player”.

:::grid {cols=2,gap=4,type=images}

:::

• Its hardware targeted video and simple interactivity for educational titles.
• As part of its deal with Nintendo, Philips published a handful of officially licensed Zelda and Mario games.

Those games were notorious — clunky controls, awkward animation, minimal oversight from Nintendo.
CD-i showed that IP licensing without platform focus doesn’t create a successful console.

What Comes Next

By the end of the 1990s, the winning formula was clearer:

• CDs (and later DVDs) reduced manufacturing cost and enabled “big” games.
• A developer-friendly platform holder could reshape third-party gravity.

The next shift wasn’t just more storage. It was consoles becoming living-room appliances — DVD players, hard drives, and online services — and the first real “service layer” taking hold.

Next: DVD Players, Modems, and Hard Drives

From there, the 16-bit era added a different kind of shift: consoles stopped being generic boxes and started becoming identities — fast, edgy, family-friendly, RPG-heavy — shaped as much by marketing and third-party ecosystems as by hardware.

Preventing Another Crash: Licensing and Curation (1980s)

Nintendo Famicom / NES — Hardware + Licensing

Nintendo’s Family Computer (Famicom) (1983 Japan) and its western cousin NES (1985 North America) revived the console market on a very different model.

::: grid {cols=2,gap=4,type=images}

:::

• Hardware:
  • 8-bit CPU with a custom PPU (Picture Processing Unit) for sprites and smooth scrolling.
  • Simple but flexible audio with multiple channels for melodies, bass lines, and noise.
• Form factor:
  • In Japan, playful red-and-white plastic with hard-wired controllers.
  • In the US, redesigned as a VCR-like front-loader to distance it from “game console” stigma.

The key innovation wasn’t just technical. It was business:

• Nintendo created a strict licensing program.
• Every licensed cartridge used a lockout chip and carried the “Seal of Quality”.
• Third parties had quotas and contract limits; unlicensed publishers were pushed out.

This tight control avoided another Atari-style crash and let Nintendo curate a library that still defines the 8-bit era:

• Super Mario Bros. redesigned platforming physics.
• The Legend of Zelda introduced battery-backed save data on a console cartridge.
• Metroid and Castlevania experimented with nonlinear world design.

Sega Master System — Technically Strong, Ecologically Weak

The Sega Master System (based on the Mark III in Japan) often looked better on paper than the NES:

• It supported more on-screen colors.
• Its Z80 CPU was familiar and tried-and-true.
• Some ports (e.g. Space Harrier, OutRun) were closer to their arcade originals.

But Sega lacked Nintendo’s licensing network, and many third-party publishers were contractually tied to releasing only on Nintendo hardware in certain markets.
Master System thrived more in Europe and Brazil, but in North America it became a niche box — proof that hardware specs alone cannot win a generation.

Consoles as Identity: The 16-bit Personality Era (late 1980s–1990s)

Sega Mega Drive / Genesis — Speed, Attitude, and “What Nintendon’t”

The Mega Drive (Japan) / Genesis (North America) pushed Sega into the lead, at least briefly.

• Hardware:
  • Motorola 68000 CPU at 7.6 MHz for the main logic.
  • Z80 coprocessor and an FM synth audio chip (Yamaha YM2612) for arcade-like sound.
• Design goal:
  • Fast scrolling, big sprites, and responsive controls — ideal for action and sports.

Marketing in the US turned it into the “cool” console:

• Slogans like “Genesis does what Nintendon’t”.
• Edgier games (blood in Mortal Kombat with a code, versus censored SNES).
• Sonic the Hedgehog as a mascot built on sheer speed.

Genesis repositioned consoles from “toys” toward teen culture.

Super Nintendo (SNES) — Enhancement Chips and RPG Heaven

Nintendo answered with the Super Famicom / SNES.

• Hardware:
  • Slower main CPU than Genesis, but a powerful PPU capable of multiple layers and the famous Mode 7 (affine transform of background planes for pseudo-3D effects).
  • High-quality sample-based audio through the SPC700 chip.

Instead of betting everything on base hardware, Nintendo embraced enhancement chips on the cartridge:

• Super FX for polygonal 3D in games like Star Fox.
• DSP and other coprocessors for fast math, rotation, scaling, and special effects.

SNES became the home of lavish 16-bit JRPGs and action adventures:

• Final Fantasy IV/VI, Chrono Trigger, Secret of Mana, A Link to the Past.
• Each one pushing story, music, and system design into territory that still feels dense today.

Early Optical Experiments (and Other Detours)

NEC PC Engine / TurboGrafx-16 — Cards and CDs

NEC’s PC Engine (TurboGrafx-16 in the West) chose a different path:

• Games came on slim HuCards (credit-card-sized cartridges).
• An optional CD-ROM add-on brought CD audio and large storage to the living room earlier than most rivals.

It shined especially in Japan with shooters and action games, and showed how optical media could power animated cutscenes and CD music even on modest hardware.

Neo Geo AES — The Luxury Arcade in a Box

SNK’s Neo Geo AES was effectively an arcade board repackaged for home.

• It used the same architecture as the Neo Geo MVS arcade system.
• Cartridges were huge and expensive; individual games could cost more than entire competing consoles.

Owning a Neo Geo meant having nearly pixel-perfect versions of Metal Slug, King of Fighters, and Samurai Shodown at home — if you could afford it.
It demonstrated a different fantasy: no compromises, just pay whatever it costs.

What Comes Next

The 16-bit era didn’t just end because polygons arrived. It ended because storage economics changed.

When CDs became cheap and huge compared to cartridges, the center of gravity shifted toward cinematic production values, lower manufacturing risk for publishers, and a new kind of platform holder.

Next: When CDs Beat Cartridges

“Console generations” are usually drawn as a simple staircase of power:
8-bit → 16-bit → 3D → HD → 4K → ray tracing.

Every few months someone declares the war over — most recently, GameStop’s Oct 25, 2025 statement joked that the console wars started with Halo: Combat Evolved exclusivity and would end once Halo: Campaign Evolved lands on PlayStation in 2026 — yet the debates never quite die.

In reality, every step hides a much messier layer:

• media choices — cartridges vs CDs vs DVDs vs SSDs
• business friction — Nintendo vs Sony vs Philips vs SEGA vs Microsoft
• ecosystem bets — tight control vs open third parties vs subscription services

This series follows that history generation by generation, but more importantly, console by console. I linger on the individual machines because each of them altered the path in some particular way.

Before Software: Dedicated TV Boxes (1970s)

The earliest home consoles weren’t “platforms” in the modern sense. They were products: sealed boxes built around a handful of built-in games.

Color TV-Game Series — Nintendo Before Famicom

Before the Famicom, Nintendo’s first home consoles were a line of dedicated machines: the Color TV-Game series.
(Technically these shipped after early first-generation systems like the Magnavox Odyssey; I’m calling them out separately because they were Nintendo’s bridge from toys to fully fledged consoles.)

These were Japan-only boxes sold between 1977 and 1980, with no interchangeable games:

• Color TV-Game 6 and 15 (1977): Pong-style tennis variations with knobs on the console itself. Players turned dials to move paddles while the console drew simple colored blocks on the TV.
• Racing 112 (1978): A tiny “cockpit” console with a steering wheel and accelerator, rendering a top-down racing lane with moving obstacles.
• Block Kuzushi (1979): Nintendo’s take on Breakout, with a paddle controller and brick-breaking gameplay baked directly into the hardware.
• Computer TV Game (1980): A dedicated version of Nintendo’s own Computer Othello arcade game, turning your TV into a digital board game surface.

::: grid {cols=3,gap=4,type=images}

:::

These machines sold a few million units in Japan and did one crucial thing:
they taught Nintendo how to design TV-connected electronics, industrial enclosures, controllers, and game logic long before the Famicom ever existed.

They were sealed boxes, but they proved something bigger: there was a home for videogames in the living room.

The Core Concept: A TV Can Be Interactive (1972)

Magnavox Odyssey — Videogames as Electronic Board Games

The Magnavox Odyssey (1972) is widely recognized as the first commercial home videogame console.

Technically, it barely resembles what came later:

• It used only analog circuitry — no CPU, no RAM.
• It output a few movable squares and a line; everything else was imagination.
• It shipped with plastic screen overlays, dice, cards, and paper money.

You didn’t “buy a game” as software; you got rule sheets telling you how to interpret those moving squares as tennis, hockey, or haunted houses.

Odyssey proved the core idea: a TV could be interactive.
Everything after it is refinement.

Swappable Games: Cartridges Create the Console Platform (late 1970s)

The first true break in home console history wasn’t 8-bit vs 16-bit. It was this:

The moment games became swappable software — and consoles became stable platforms.

That shift created the third-party business, the boom… and the first crash.

Fairchild Channel F — Swappable Software

The Fairchild Channel F was the first console with ROM cartridges you could plug in and out.

• Built around a Fairchild F8 CPU, it split games off from the hardware.
• Cartridges allowed developers to ship new experiences without changing the console itself.

Its library was modest and it lost the commercial race, but it introduced an architecture that every later console inherited:
a stable hardware base with evolving software on top.

Atari 2600 — The Rise and the Burnout

The Atari 2600 (1977) turned that architecture into a cultural phenomenon.

• Hardware: a tiny 8-bit CPU, 128 bytes of RAM, and a video chip that programmers had to “race” line-by-line to draw anything at all.
• Library: Space Invaders, Pitfall!, Adventure, early sports games — all crammed into cartridges as small as 2–4 KB.
• Business model: anyone could make a cartridge for it.

That openness created the first third-party wave (Activision comes from ex-Atari developers), but also a flood of low-quality games.
By 1983, the market was overloaded with bad cartridges; retailers dumped stock and pulled back shelf space. The North American videogame crash followed.

Atari 2600 proved two things at once:

1. A home console business could be huge.
2. Without control over software quality, it could collapse just as fast.

What Comes Next

The crash didn’t kill home consoles — it forced them to grow up.

Nintendo’s next move wasn’t primarily technical. It was managerial: licensing, lockout chips, and quality control — a new rulebook designed to prevent the platform from destroying itself.

Next: Curation, Characters, and Enhancement Chips

此行登車在凌晨，僅得幾小時睡眠，醒時天尚未明。周遭靜極，只聽得車輪與軌道低聲磨合，似水波拍岸，又似夢語餘韻。人未全醒，心卻已有了些動靜。披衣起身，倚坐靠窗，杯中水未冷透，拂面風涼而不寒，忽地便覺這段路程，不必抵達也好。

車廂不喧，有如舊時客館。身邊並無同行之人，也不覺孤單。這樣的早晨，自有其分寸之美。行李擱在腳邊，簾未全卷，光線不甚明亮。山間的天色本就如此——未雨，卻濕；未晴，卻亮；不言清朗，亦非昏沈。正適合想些不必說出口的事。

腦中浮起一些舊影，也無甚特定，只是在這樣緩慢的節奏裡，被悄悄喚起。童年時常隨家人南來北往，十數小時車程也無怨，鋪上翻滾、瓜果分食，是路途中的熱鬧，也是年歲裡最長情的陪伴。那時的自己，困極了也不肯獨自入眠，總要黏在熟悉人身旁才安心。

如今再行車途，所伴之人已不在側。只是坐於車廂一隅，竟覺好似與那年小小的自己重逢。也無甚話要說，只偶爾低頭看看手中茶水，便生出些不知從何而來的溫柔與靜氣。

人總是這樣，走得太快了，反倒忘了許多走過的路。只有在這樣不催不趕的晨光裡，在這樣霧氣浮動、不甚清明的天氣裡，記憶才會不動聲色地浮上來，像水中碎光，一閃便散，卻又教人久久難忘。

並非刻意去追尋什麼，也無意埋怨什麼。只是偶爾想起那些曾在我生命裡留下溫意的人事物，有些已走遠，有些未曾靠近。當時不覺，如今回看，種種隱約，似也另有深意。那年歲裡藏得住的沉默，到了今日，也不過是一片柔軟的懷念罷了。

行至一地，終會離去。心中所念，也不過如此。無需言語，無需回聲，只是知道，曾有那麼一段路，於無聲中照亮過我，於無形中輕輕擔過一些原本無需他人承受的重擔。這樣的記憶，是溫的，不熱烈，卻始終不冷。

車仍緩行，霧未散盡。晨光從薄雲中滲出些許，映在牆面與指背上，並不耀眼，卻足夠溫暖。許多事，其實都無需說得太明。心中若有光，便自知去處；若有念，風也會懂。

窗外仍是層疊山影，時明時暗。這一路，或許無人同坐，但也無妨。偶爾靜靜地與某段舊年並肩而行，已是難得。

終點會到，但此刻不急。世事紛擾，總需一段這樣安靜的時光，來與從前的自己細細對望。而有幸，我正在途中。

On paper, the Nintendo Switch 2 is exactly what the original system’s fans spent eight years asking for: a larger 7.9-inch 1080p HDR screen, a much faster custom Nvidia SoC, 12 GB of LPDDR5X memory, 256 GB of UFS internal storage, and 4K HDR output in docked mode.

In the market, it is an overwhelming success. Launch sales broke internal records and set new industry benchmarks, with millions of units sold in the first few days. For Nintendo as a business, the hybrid idea has never looked stronger.

Yet as a hardware generation, the Switch 2 remains constrained by the same triangle that shaped the original Switch — heat, power, and battery life — while sharing a software era with the PS5 and Xbox Series consoles. The result is commercially formidable and technically ambitious, but less decisive as a generational break than its specifications first suggest.

What the Switch 2 Actually Is

The Switch 2 keeps the hybrid design almost untouched: a tablet-style console with detachable Joy-Con 2 controllers, a dock for TV play, and a single OS that spans handheld and living-room modes. The underlying silicon, however, is a clear leap over the Tegra X1 in the original Switch.

At the center of the system is a custom Nvidia Tegra T239 (“Drake”) SoC, with a cluster of modern ARM CPU cores and an Ampere-based GPU that finally belongs in the same architectural family as contemporary PCs and consoles. This is paired with 12 GB of LPDDR5X memory on a 128-bit bus, giving both higher bandwidth and more headroom for modern engines. Internal storage jumps to 256 GB of UFS, a tacit admission that the 32 GB of the original Switch was never realistic for current-generation games. For expansion, Nintendo moves to microSD Express for faster game loading.

The display steps up to a 7.9-inch 1920×1080 LCD with HDR support and variable refresh up to 120 Hz in handheld and tabletop modes. Docked, the console can output up to 4K/60 with HDR over HDMI, and it adds high-refresh 1080p and 1440p modes for more PC-like monitors.

The Joy-Con 2 controllers shift to a slightly taller shell and magnetically attach to the sides with tighter tolerances than before. ZL/ZR buttons grow larger, and a dedicated “C” button appears, tied to Nintendo’s new GameChat feature. Rotated and placed on a desk, each Joy-Con can act like a hybrid between a small gamepad and a pointing device. Hopes for Hall-effect sticks never materialize; instead, Nintendo advertises improved durability within a familiar analog-stick design.

On the social side, the platform pushes further than the original Switch ever did. GameChat layers voice communication, screen sharing, and optional camera input directly into the OS. The built-in microphone and accessory support for headsets or webcams make the system feel less like a sealed toy and more like a socially aware terminal. Even the Mii system has been modernized, moving away from rigid gender selection toward style-based customization that reflects contemporary expectations of identity.

On spec sheets, the Switch 2 looks like a clean, forward-looking hybrid platform. Once it is placed in a real thermal envelope and plugged into real development pipelines, the compromises become much harder to ignore.

Where the Generational Leap Narrows

Commercial momentum and a convincing generational transition are not the same achievement. As a next-generation system, the Switch 2 must negotiate technical, ergonomic, and ecosystem limits that keep it from fully expressing the capabilities suggested by its silicon.

Its shortcomings can be understood across three intertwined dimensions: performance constraints, visual presentation limits, and ecosystem stagnation.

Performance Trap: The Inescapable Triangle of Heat, Power, and Battery Life

The Switch 2’s hybrid nature forces mutually incompatible requirements onto its architecture.

In docked mode, the T239’s Ampere GPU can be clocked high enough to flirt with “last-gen home console” territory. With reconstruction techniques and hardware ray tracing in the mix, it finally has enough headroom to run modern engines at respectable resolutions and frame rates, especially when targeting 4K output via upscaling. On a living-room screen, the system can present itself as a credible, if modest, competitor.

In handheld mode, reality takes over. The 7.9-inch 1080p HDR panel, the 120 Hz VRR capability, Wi-Fi 6, 12 GB of high-speed DRAM, and a relatively power-hungry GPU all share a fixed battery budget. To maintain tolerable battery life and safe thermals, the console must cut clocks, lower power targets, and trim visual settings. A demanding title that tries to push the hardware’s theoretical limits will rapidly collide with both the thermal design and the battery capacity.

The user-facing pattern is familiar. Docked, Switch 2 titles chase impressive targets—4K output, high frame rates, ray-traced reflections—but only within carefully curated scenarios. Undocked, the same games often fall back to reduced internal resolution, simplified effects, and more conservative frame pacing, even when the marketing headline still reads “1080p portable play.”

Upscaling and reconstruction techniques can do a great deal, and in the best cases the results are surprisingly clean. But they cannot override the physics of a thin handheld chassis and a modest battery pack. The silicon is capable of more than the form factor can sustain continuously. So long as the console must be both a couch system and a portable, it cannot compete on exactly the same terms as dedicated living-room hardware.

Visual Ceiling: A Screen That Cannot Fully Express the Leap

On the surface, the Switch 2’s display seems like the long-awaited fix: a step from 720p to 1080p, SDR to HDR, and a 60 Hz panel to one that can handle 120 Hz VRR. For a portable system, that sounds almost excessive.

Yet this is also where the generational leap stops short of being genuinely transformative. The panel is still LCD rather than OLED, which means black levels, contrast, and HDR “pop” have hard ceilings. HDR support exists, but within the constraints of a mobile display whose brightness cannot simply be driven to desktop monitor levels without destroying battery life. The higher resolution is welcome, but its impact is softened whenever games render below native internally and rely on reconstruction or upscaling to reach 1080p.

The overall impression is of a carefully optimized evolution. Text and HUD elements are sharper; textures and fine detail read more clearly than on the original Switch; shimmering and aliasing are reduced when developers lean into modern anti-aliasing. But these changes feel like an accumulation of refinements rather than a qualitative redefinition of how games look and feel on a portable screen. Across a table, a Switch 2 often resembles a very polished original Switch: recognizably the same idea, only cleaner.

Internally, the SoC belongs to a new era. Externally, the main display still behaves like a conservative, power-aware update to a 2017 philosophy. The emotional shock of a generational transition—the sense that old franchises are being reframed through a radically upgraded image pipeline—is dulled.

Development Stall: A Next-Gen Console Without Next-Gen Development Cycles

The deepest structural problem is neither in silicon nor in glass, but in time.

During the early life of any console, developers quietly decide how “real” its generation will be. They can rebuild engines, streaming systems, and asset pipelines around the new baseline, or they can treat the newcomer as a slightly higher configuration in an existing multi-platform roster. For Switch 2, economic gravity pulled strongly toward the second option.

Production pipelines had been tuned to the original Switch for years. Geometry budgets, texture resolutions, simulation complexity, and loading behavior were all shaped by its limited CPU, memory, and storage profile. When Switch 2 appeared, its predecessor’s enormous install base did not vanish. The rational strategy was to keep building for the original target while layering Switch 2 modes on top: higher resolutions, cleaner textures, more stable frame rates, and occasional extra effects.

This decision has consequences that extend across the generation. Switch 2 struggles to carve out a distinct “native” identity in its formative years, because the majority of its library behaves like “upgraded Switch games” rather than software fundamentally conceived for a new baseline. At the same time, the older hardware remains the gravitational center of Nintendo’s third-party ecosystem; many cross-platform titles are simply not allowed to assume that the Switch 2 profile is the minimum.

In other words, the console’s generational leap is partially sacrificed to preserve continuity with the economic reality of its predecessor. By the time this inertia begins to fade, the crucial early window in which a platform establishes its identity will already have passed.

Dock, Docks, and the Fragile 4K Dream

The dock is once again where Nintendo tries to convince the world that the Switch belongs in the same living-room space as Sony and Microsoft.

The official Switch 2 dock is clearly more serious than the original. It formally supports 4K/60 HDR output, wired networking, and offers modes targeted at 1080p and 1440p high-refresh displays. It feels less like a plastic HDMI breakout and more like a deliberate bridge between the handheld and a modern TV or monitor.

Underneath, however, the ecosystem around the dock is fragile and tightly managed. The console relies on a proprietary link and a stricter power-delivery handshake than before, which immediately broke compatibility with a range of existing third-party docks and USB-C hubs. Hardware that had worked perfectly well with the original Switch suddenly fell back to basic charging or lost video output altogether when paired with the new system.

Firmware updates have already demonstrated how precarious this layer can be. A single system update was enough to disable video on certain unofficial docks until their manufacturers pushed out their own patches. Enthusiast users have learned that Switch 2’s TV-facing side is not a stable, standards-driven environment but a moving target, defined entirely by Nintendo’s next system software revision.

For a console that sells itself as a 4K HDR living-room endpoint, this tight coupling between dock, firmware, and video output feels out of step with the broader HDMI and USB-C ecosystem.

A Beautiful, Constrained Social Machine

If one ignores the thermal envelope and dock politics, the Switch 2 is in many ways a beautifully constructed social machine.

GameChat is the most obvious symbol of this direction. Voice chat, screen sharing, and optional camera input are integrated into the OS rather than left to external devices. The built-in microphone on the tablet, the accessory options for headsets and webcams, and the ability to share a handheld or docked session directly from the console all mark a clear departure from the isolated feel of the original Switch. It is easier to imagine the Switch 2 at the center of a small social group, not just as a local multiplayer device but as a node in an online conversation.

The new Mii system contributes in its own quiet way. By moving away from binary gender selection and toward style-driven customization, Nintendo updates a relic from the Wii era without discarding its charm. Miis remain unmistakably Miis, but the editor now reflects a more flexible understanding of identity and presentation.

Backwards compatibility is handled with a pragmatic mix of continuity and segmentation. The bulk of the original Switch library, both physical and digital, runs on the new hardware with minimal friction. Certain high-profile titles receive dedicated enhancements or full “Switch 2 editions” with upgraded assets and features, creating a soft dividing line between generations without a hard break. Subscription-based access to classic systems continues and expands, pulling older hardware histories into the Switch 2’s orbit.

At the same time, the account and enforcement layer becomes more central. Terms of service, regional policies, and ban mechanisms make clear that this is not meant to be a general-purpose computing device. It is a curated social appliance whose boundaries are tightly defined and actively policed.

Between Generations: Where Switch 2 Actually Lives

By calendar and catalog, it is straightforward to label the Switch 2 as a “tenth-generation” console. In future timelines, it will sit opposite the PS5 and Xbox Series hardware, filling Nintendo’s box for that row of history.

Experience tells a more ambiguous story. On one axis, Switch 2 is constantly measured against PS5 and Series X/S for multi-platform releases, and here the gap in raw power and thermal headroom is impossible to conceal. Even with reconstruction, dynamic resolution, and careful engine work, the system cannot reliably reach comparable resolutions and frame rates, particularly when untethered from the dock.

On another axis, the console is chained to its own past. The original Switch’s huge install base and rich software catalog exert a gravitational pull on both publishers and players. “Switch” becomes a single amorphous platform in public perception; Switch 2 is treated less as a break and more as the premium configuration in that ecosystem.

This dual constraint explains why the Switch 2 can triumph as a product while still feeling hesitant as a generational break. Its thermal and battery envelope limits how fully the SoC can express itself in handheld use. Its conservative display softens the visual drama of a major internal upgrade. And a software ecosystem shaped by compatibility and economics takes time to gather around a new baseline.

Nintendo’s first-party teams will, as always, find ways to make the hardware look better than it has any right to. They will design worlds around the Switch 2’s strengths instead of treating it as a downscaled target, and their work will define the system in collective memory. But even their best efforts cannot completely erase the underlying compromises.

A Generation That Ends the Previous One, Instead of Starting the Next

The Nintendo Switch 2 is a paradox: a successful, forward-looking hybrid whose strongest quality — continuity — is also what makes its generational identity difficult to see clearly.

As a physical object, it is easy to like. The tablet is more comfortable, the Joy-Con connections are better, the screen is sharper and supports HDR, and the social layer finally feels native rather than bolted on. As a business, it refines a model that has already transformed Nintendo’s fortunes and will likely do so again.

As a generational statement, it hesitates. It does not join the PS5 and Xbox Series on the same performance frontier, nor can it escape the gravitational pull of the original Switch’s lower baseline. It instead occupies a liminal space: the perfected form of a 2017 design decision, powerful enough to extend that idea but not quite willing to redraw its boundaries.

In the long view of console history, the Switch 2 looks less like the opening of a new chapter and more like the definitive closing paragraph of the previous one.

The so-called advantages of vision-only perception — lower cost and algorithmic maturity — are becoming less compelling in today’s highly competitive landscape. TOF and FMCW LiDAR technologies are rapidly dropping in price; even a relatively ordinary project I’ve worked on recently has adopted a solid-state TOF solution. Meanwhile, sensor fusion algorithms are inevitably maturing as well.

The depth maps generated by vision-only systems are essentially pseudo-3D representations. When the technology is immature, depth estimation tends to suffer from errors and is easily deceived by environmental conditions. In contrast, LiDAR can offer millimeter-level — and with FMCW, even sub-millimeter — precision, along with significantly improved resistance to interference.

Moreover, using neural networks to compute depth and perform object detection demands substantial computational power. LiDAR, on the other hand, directly outputs 3D data, greatly reducing the processing burden.

Vision-only perception systems are also highly dependent on lighting conditions and scene texture complexity. LiDAR remains stable under backlight, low-light, and textureless scenarios, and continues to function reliably even in adverse weather such as rain and fog.

The lack of safety redundancy in vision-only approaches makes them unreliable for use in such challenging conditions — and realistically, we can’t expect users to simply avoid driving at night or in bad weather.

The future of autonomous driving will most likely rely on multi-sensor fusion: a form of generalized perception that balances cost and robustness by drawing on the strengths of several modalities.

The goal should not be to vindicate vision-only perception as a principle, but to build heterogeneous systems in which vision provides the backbone and LiDAR supplies redundancy and precision where they matter most.

For high-level autonomous driving (L4 and above), TOF — and perhaps especially FMCW LiDAR — is therefore likely to become increasingly common.

Specialized machines will continue to do most individual jobs better. Yet the sensors, actuators, control systems, and commercial lessons developed for those machines can also become the modular foundations of more general-purpose platforms. Progress may therefore move in both directions: specialized robotics lending maturity to humanoids, and humanoid research creating new demand for better specialized components.

Specialized systems are unlikely to disappear, especially where performance, reliability, or cost matters more than adaptability. In that respect, the relationship between humanoid and specialized robots resembles the relationship between AGI and vertical AI: breadth is valuable, but it does not make depth obsolete.

I remain skeptical of both humanoid robots and AGI, particularly of the assumption that generality will arrive at an acceptable cost. A machine capable of adapting to nearly every human environment may be possible long before it is economical.

Even so, the possibility that these two lines of research might one day meet — and produce something recognizably like embodied AGI — remains difficult not to contemplate.

To understand how consoles preserved — or abandoned — earlier software is therefore to understand how the medium negotiates with its own history.

The Era of Hardware Inheritance: Hiding the Previous Console Inside the New One

In the early generations of game hardware, software was tightly coupled to the specific quirks of the underlying circuitry. CPU timing, bus behavior, scanline progression, undocumented registers, race conditions, and even hardware bugs were part of the expected execution model. Many games depended on these characteristics in subtle ways. In such an environment, the most robust way to ensure compatibility was simply to embed the previous generation’s hardware inside the new system.

Nintendo’s early handheld line exemplifies this logic. The Game Boy Color contained a Z80-like core compatible with the original Game Boy. The Game Boy Advance went even further: beyond its ARM7 processor, it preserved a full set of Game Boy Color graphics registers, pixel timing units, and DMA behavior. When running GBC software, the GBA did not “simulate” the older handheld—it effectively transformed into one. This produced impeccable compatibility at the cost of additional silicon area, power consumption, and board complexity.

Such designs were nearly flawless in behavior but carried high physical and economic costs. Once miniaturization and sleek industrial design became dominant goals—most visibly in the Game Boy Micro—it became impossible to retain the older hardware blocks. Backward compatibility became incompatible with the physical constraints of the product.

The Hybrid Era: Partial Hardware Retention and Software Compensation

As console architectures grew more sophisticated, maintaining full hardware inheritance became increasingly impractical. Manufacturers turned to hybrid models: retain only the most essential parts of the previous system’s architecture, and reconstitute the rest through software layers or lightweight auxiliary circuitry.

The PlayStation 2 is a canonical example. Early PS2 units included the full PS1 CPU and critical graphics logic, enabling near-perfect compatibility. Later slim models removed some dedicated hardware blocks, relying instead on software implementations for portions of PS1 behavior. Compatibility remained high, but subtle divergences emerged in edge cases.

Nintendo’s Wii adopted a similar strategy. It contained the logic necessary to become, in effect, a GameCube, dropping into a hardware “GC mode” that bypassed much of the Wii subsystems. Stability and fidelity were excellent, but the approach still required the physical footprint of legacy logic.

Hybrid compatibility represented a transitional phase: no longer carrying the entire past, but still unable to completely shed hardware obligations.

The Software Emulation and Remaster Era: Compatibility as an Engineering Service

With substantial increases in CPU and GPU performance, consoles began shifting fully toward software emulation. Instead of physically rebuilding old hardware, engineers rebuilt its behavior: instruction sets, timing, rasterization pipelines, audio envelopes, and even precise hardware quirks. Compatibility ceased to be a hardware commitment and became a software engineering project.

This shift defined much of the modern landscape. PlayStation 3 later revisions, Xbox One’s backward compatibility program, and numerous digital re-releases rely on emulation frameworks. Software emulation allowed old games not only to run, but to benefit from enhancements: save states, rewind features, resolution boosts, anti-aliasing, and CRT shaders—all impossible on original hardware.

At the same time, the dominance of software compatibility paved the way for a parallel phenomenon: remasters and remakes. Once games no longer depended on legacy hardware, developers could rebuild them directly for modern systems—sometimes using original engines wrapped in emulation, sometimes recreating them from scratch. Thus “compatibility” bifurcated into two paths: emulation that recreates the original experience, and remasters/remakes that reinterpret it.

This dual pathway defines the contemporary understanding of backward compatibility: a mix of technical preservation and commercial reinvention.

Handheld Evolution and the Rise of Virtualized Execution Environments

Handheld systems followed the same general trajectory but manifested it in a more compressed and extreme form due to severe constraints on size, thermals, and battery life.

Nintendo’s GB/GBC/GBA lineage originally relied on hardware inheritance, but the need for aggressive miniaturization ultimately forced the Game Boy Micro to abandon all pre-GBC compatibility. The Nintendo DS briefly reintroduced physical backward compatibility via a dual-slot design, only for the 3DS to consolidate everything into a single SoC and rely on a hardware-level “DS mode” rather than a discrete cartridge subsystem. By the time the Switch arrived, physical backward compatibility vanished entirely, replaced by emulation-driven digital libraries.

Sony’s handheld strategy took a different but equally illustrative path. The PSP contained no PS1 hardware; it relied entirely on an official emulator bundled uniquely with each PS1 digital release. The PlayStation Vita refined this approach, implementing a virtualized execution environment capable of running PSP titles on a fundamentally different architecture. Rather than inheriting hardware, the Vita hosted reconstructed software environments—an early form of handheld OS-level virtualization.

This evolution marks a turning point: handheld backward compatibility no longer depended on the past’s physical architecture, but on the platform holder’s willingness to maintain digital distribution and software layers. Compatibility became a platform service rather than a hardware trait.

The Cost of Compatibility: Why 3DS Backward Support Became a Security Vulnerability

Backward compatibility also carries risks. The Nintendo 3DS illustrates how compatibility can inadvertently preserve not only legacy functionality but legacy vulnerabilities.

To support the Nintendo DS library, the 3DS included a dedicated DS execution mode. When engaged, this mode replicated the DS security model, filesystem structure, and peripheral behavior. Unfortunately, it also replicated the DS’s weaker security boundaries. Early 3DS exploits began precisely in this DS mode: attackers leveraged legacy flaws in image parsers, savefile handlers, and cartridge command paths to gain arbitrary code execution. With carefully constructed payload chains, they escalated from DS-mode control into the full 3DS environment.

In this sense, backward compatibility became a historical liability. It preserved not only software but the vulnerabilities of a previous generation. This phenomenon notably influenced Nintendo’s later decision to avoid hardware-level backward compatibility in the Switch, preferring sandboxed emulation layers that isolate old logic rather than replicate it.

Backward compatibility, therefore, is never purely “good.” It is always a trade-off between preservation and inherited fragility.

Emulator Culture: The Unofficial Foundation of Modern Compatibility

Behind the industry’s evolution lies another longstanding force: emulator culture. Independent emulator developers spent decades reconstructing undocumented hardware behavior through painstaking reverse engineering—measuring timing cycles, testing register responses, tracing bus activity, and identifying undocumented behaviors. Much of what is publicly known about older systems comes from this work, not from official documentation.

Over time, emulator research effectively became the backbone of modern game preservation. Many official compatibility solutions—including those in the Switch Online service, PlayStation Classics, and Xbox backward compatibility—are built on methodologies pioneered by the emulator community: JIT translation, precise timing models, API redirection, shader pipeline recreation, and more.

Emulation does more than allow old games to run. It preserves behavior, makes that behavior open to analysis, and turns ageing hardware into systems that can be reproduced in software. Emulators have therefore become a cultural and historical infrastructure as important as any official backward-compatibility program.

Looking Ahead: The Future of Backward Compatibility

Backward compatibility has traced a winding path across console history. It began with literal hardware inheritance: placing yesterday’s chips inside tomorrow’s devices. It then transitioned to hybrid modes where only fragments of the past survived on silicon. Eventually, it became a software discipline—an engineering practice rather than a hardware obligation—accompanied by remasters that reinterpreted classics for new generations.

Handheld systems, from GBA to PSP and PSV, mirrored this evolution in more extreme forms, moving rapidly from inheritance to virtualization. In systems like the 3DS, compatibility even revealed its darker side by inheriting old vulnerabilities. And throughout this process, emulator culture provided the conceptual and technical framework that made modern compatibility—official and unofficial—possible.

Each compatible console carries the past differently: as an extra chip, a hidden execution mode, an emulator, or sometimes only a listing in a digital store. As architectures converge and physical media recede, the engineering may become easier while the act of preservation becomes more dependent on corporate choice. The question will no longer be only whether an old game can run, but whether its platform holder still wishes to let it.

Among the many PS3 revisions, only the CECHA (60GB, Japan/US) and CECHB (20GB) models contained an entire PlayStation 2 system integrated on the motherboard — including the Emotion Engine CPU and Graphics Synthesizer GPU.
This meant true, hardware-level backward compatibility, allowing the system to run games from all three generations — PS1, PS2, and PS3 — with perfect accuracy.

It was the ultimate “all-in-one” PlayStation.
A vision of seamless continuity between generations — and, in retrospect, one of Sony’s most daring hardware decisions.

The Dream and Its Flaws

This ambition, however, came at a cost.
The CECHA suffered from an early thermal design flaw, which made it prone to the infamous Yellow Light of Death (YLOD). Combined with its glossy piano-black finish and limited production run, pristine units with original packaging and documentation are now exceedingly rare.
Collectors today search for untouched systems — with uncracked warranty seals and no signs of reflow repairs — like museum artifacts of a lost technological optimism.

The Controller Lineage

To fully appreciate the CECHA, it helps to look back at the evolution of the PlayStation controller itself — a lineage that mirrors the console’s history of refinement.

During the PS1 era, three major controllers were released:

1. Original Controller (Model H):
   A minimalist design without analog sticks — simple, symmetrical, almost toy-like by modern standards.

2. Dual Analog Controller:
   Introduced the now-iconic dual-stick layout, giving players precise control over movement and camera — a major leap in input expressiveness.

3. DualShock Controller:
   Added vibration feedback, merging tactile sensation with gameplay — an innovation that set the template for all future PlayStation controllers.

The PS2 continued this legacy with the DualShock 2, refining the analog sensitivity of buttons and sticks while retaining the same ergonomic silhouette.

The Silent Gap — SIXAXIS and the Patent War

Yet by the time the PlayStation 3 launched in 2006, something curious had changed.
The launch units — including the CECHA — shipped with SIXAXIS controllers, which lacked vibration entirely.

The reason was not technical but legal.
A company named Immersion had sued Sony, claiming that the vibration mechanisms in the DualShock 1 and DualShock 2 infringed upon its patents.
As a result, Sony temporarily removed vibration feedback from the controller lineup, introducing motion sensing (the “six axes” of movement) as a partial substitute.

It was an odd compromise — and players noticed.
The absence of tactile feedback left early PS3 titles feeling somewhat detached, sterile even, especially for players accustomed to the visceral response of a rumbling DualShock.

Eventually, Sony and Immersion reached a settlement.
By 2007, vibration made its return in the form of the DualShock 3, combining SIXAXIS motion control with the familiar tactile feedback players had missed.
It restored what many felt was the soul of the PlayStation experience.

The End of an Era

The CECHA thus represents a fascinating intersection in gaming history:
a console where the old and new coexisted — full hardware backward compatibility, legacy controller design, and a vision of total integration that would never be repeated.

Later PS3 revisions gradually removed PS2 hardware, replacing it with limited software emulation to reduce cost and heat output.
Backward compatibility became partial, and eventually disappeared altogether.

In hindsight, the CECHA was the last PlayStation built around a genuinely universal idea of the platform. Its ambition was unsustainable, but the premise remains compelling: technological progress did not yet have to mean abandoning the past.

For collectors and enthusiasts, the CECHA remains a physical reminder of a time when “compatibility” meant continuity, and when a single machine could still carry the memory of three generations.

When Nintendo launched the Switch in 2017, the home console landscape was already split in two. High-performance boxes under the TV were chasing 4K and HDR, while smartphones and tablets had eaten most of the handheld market. The Wii U had died in that no man’s land between the two: not powerful enough to stand alongside PS4 and Xbox One, and not portable or elegant enough to be a true handheld.

The Switch responded with a blunt, almost radical proposition: stop choosing. Instead of iterating on “home console” and “handheld” as separate product lines, Nintendo fused them into a single object — a tablet-like core that could be both, depending on where you put it and how you held it.

The commercial result is already carved into sales charts. The Switch has climbed into the same altitude as the DS and PS2 era, and its software library has sprawled across most of a decade. For a machine built around a mobile SoC and a 720p screen, its influence and lifespan are absurdly out of proportion to its raw specs.

Cutting the Safety Net: Deliberately Breaking Backward Compatibility

From a hardware romantic’s perspective, one of the Switch’s most controversial decisions is how aggressively it cuts itself off from the Wii / Wii U / DS lineage. There is no disc drive, no second cartridge slot, no clever adapter that allows you to slide in 3DS cards or Wii discs. Past generations are treated as past generations.

Under the hood, this is more than just a business choice. The Switch abandons the PowerPC architecture that powered the GameCube, Wii and Wii U, and instead moves to an Nvidia Tegra SoC with ARM CPU cores and a Maxwell-family GPU. Instruction set, graphics pipeline, operating system — almost everything fundamental is different.

Real, low-level hardware backward compatibility would have meant one of two things: bolting an entire legacy subsystem onto the board, or building a labyrinth of emulation layers into the OS. Both would have added cost, complexity, and constraints to a device whose central promise is “simple, small, hybrid”.

Nintendo chose the clean break. The upside is conceptual purity. The Switch can be defined entirely on its own terms, without having to drag a decade of design decisions behind it. The downside is obvious to anyone staring at a box of Wii and 3DS software: there is no direct path forward. Those games live and die with their original hardware, unless and until Nintendo resells them via ports, remasters, or subscription services.

In effect, backward compatibility was reclassified from “physical feature” to “content strategy”. Instead of keeping an old cartridge slot alive, Nintendo selectively rebuilds a classic library on the new platform, often with updated pricing and new digital ownership rules. The Switch’s success suggests that mainstream players were willing to accept that trade, even if preservationists and collectors are less forgiving.

One SoC, Two Personas: Home Console and Handheld in One Shell

At a glance, the Switch hardware is almost comically simple: a 6.2-inch tablet-like main unit with rails on the sides, a pair of detachable Joy-Con, and a dock that is little more than a breakout box for HDMI and power delivery. The complexity is in how this small system behaves like two different machines.

Inside the main unit, the Tegra SoC runs in two broad regimes. In handheld mode, it targets the built-in 720p display with conservative clocks and power limits, pushing just enough pixels for a comfortable image while keeping battery life and thermals under control. Slot the console into the dock and the system is allowed to draw more power, ramp clocks, and output 1080p to the TV.

From the user’s standpoint, the transition is absurdly low-friction: pull the console out of the dock and keep playing in your hands; drop it back in and the TV wakes up. Underneath that single motion lies a full perspective shift — the same game logic and save data, but a different performance envelope and presentation.

On paper, the Switch is permanently outgunned by its contemporaries: 4 GB of RAM, a mobile GPU, and modest bandwidth are not the ingredients from which spec sheets are made. But as a lived object, it optimized for something else. Frame rates become “good enough”; the important part is that the console is always where you are. The hybrid concept is not a gimmick bolted on top; it is the baseline assumption.

This also explains why the Switch feels strangely resilient even as a successor arrives. As long as the little tablet and its dock are there, charged and within reach, it keeps functioning as “the machine you actually play on”, regardless of how many teraflops the newer box under the TV advertises.

Joy-Con and Sensor Overload: A Controller as Experiment Platform

The Switch story is not only about the main unit. Much of its character resides in the Joy-Con — two tiny controllers that try to be many things at once.

Each Joy-Con houses an accelerometer and gyroscope, an HD rumble motor, a full set of buttons, and a tiny battery. The right Joy-Con adds an IR motion camera; both include NFC for Amiibo. It is an extravagant amount of hardware squeezed into two pastel-colored bricks.

This layout allows at least three distinct modes of use. Clipped to the sides of the console, the Joy-Con form something that resembles a handheld; slotted into a grip, they mimic a traditional gamepad; pulled apart and rotated, they become two minimalist controllers for impromptu multiplayer sessions. Certain games push the hardware further, treating the Joy-Con as motion-sensing wands or fitness sensors.

Where the Wii Remote invited exaggerated arm swings and theatrical gestures, the Joy-Con leans toward subtler motions and haptic illusions: the suggestion of ice cubes rattling in a glass, the sensation of a fishing line tightening, the distinction between a gentle twist and a sharp flick. It is an attempt to make motion control feel less like a party trick and more like another dimension of input.

The irony is that the Joy-Con is just as famous for drift and repairs as it is for clever design. Complex sensors and miniaturized analog sticks inside a small shell, used across years of heavy, portable play, is a recipe for wear. In hindsight, Nintendo overreached on mechanics and underbuilt for durability. The idea of the Joy-Con is fantastic; the execution is fragile.

Labo: Cardboard Shells and a One-Off Education Experiment

Nintendo Labo may be the most unusual branch of the Switch ecosystem. It takes all those latent Joy-Con sensors and asks a peculiar question: what if children could build the hardware themselves, out of cardboard?

The kits include pre-cut cardboard sheets and a game card. Folding, slotting, and taping turns flat pieces into pianos, fishing rods, handlebars, houses, and even a wearable robot backpack. The Joy-Con slide into these structures, their sensors suddenly repurposed. The IR camera in the right Joy-Con “reads” which piano key you pressed; the accelerometer and gyroscope watch how you tilt the handlebars; the rumble motors bring fishing lures and robot footsteps to life.

Labo is both a toy and a lesson in systems, disguised as craft. The on-screen building instructions guide the assembly process in slow, clear animations; once everything works, you can peek behind the curtain and see sensor data visualized, buttons and levers mapped to virtual events. Toy-Con Garage extends this into a node-based logic editor, letting kids wire “when this moves, do that” behaviors without touching a single line of code.

It is brilliant, and also almost destined to be transient. Cardboard bends, tears, absorbs moisture, and eats closet space. Software support is finite by design. Labo burned brightly for a moment, sold its million-something kits, then faded into the background as the market shifted back to more conventional software. In a few decades, intact Labo sets and still-working cardboard pianos will probably feel like artefacts from a brief, optimistic moment when a company decided that a game console could double as a STEM classroom.

Pandemic Years and an Extended Generation

The other force that stretched the Switch’s lifetime had nothing to do with clock speeds or software roadmaps. It was the global shock of COVID-19.

When lockdowns and social distancing hit, living rooms and bedrooms suddenly had to absorb work, school, social life, and entertainment all at once. Demand for game hardware and software spiked; supply chains, meanwhile, were under stress. The Switch found itself in exactly the right (or wrong) place: a relatively affordable console, easy to share within a household, capable of both solitary and communal play, and portable enough to move between rooms and roles.

Animal Crossing: New Horizons became the emblem of that era, a small-town life simulator turned substitute for commutes, coffee shop conversations, and casual social rituals. Ring Fit Adventure, originally a quirky fitness experiment, transformed into a lifeline for people trying to stay active in one room apartments.

In that context, Nintendo held its hand. Rather than rushing out a generational replacement, it iterated quietly with minor hardware revisions: better battery life here, an OLED display there, small board shrinks and silent internal updates. The Switch stayed “the current Nintendo system” for much longer than a classic five- or six-year cycle would suggest.

By the time a clear successor emerged, the original Switch had become more than a single product cycle. It was the console that bridged the pre-pandemic and post-pandemic worlds; the machine that children grew up with in isolation and then carried with them back onto buses, planes, and campuses. It anchored an unusually long “generation” simply by refusing to exit the stage.

Hacking and the Grey Zone: From Softmods to Hardware Mods and Flashcarts

In parallel with its official story, the Switch accumulated a dense, messy history of hacking, modding, and legal skirmishes.

Early, unpatched Switch units shipped with a vulnerability in the Tegra boot ROM. By forcing the console into a hidden recovery mode and sending a carefully crafted payload over USB, it was possible to seize control of the system before Nintendo’s firmware booted. The clever part was how players entered that mode: by shorting specific pins inside the right Joy-Con rail and holding a button combination, they could trigger the exploit without any specialized tools.

RCM jigs — tiny plastic or 3D-printed clips that shorted those pins reliably — became ubiquitous. A paperclip and a USB-C cable were enough to open the door to custom bootloaders, homebrew, and full-blown custom firmware. This “softmod era” defined hacking for early Switch owners: no soldering iron required, but the threat of bans and the ethical questions around backups and piracy hung over everything.

Nintendo’s response was predictable. Newer hardware revisions patched the boot ROM flaw, cutting off the RCM exploit for good. Owners of these later units now needed a more invasive route. The hacking scene moved down a level, onto the motherboard itself.

Enter hardware mods built around microcontrollers such as the RP2040. These small boards, soldered to test pads and signal lines on the Switch mainboard, behave like parasitic co-processors. By glitching voltage, clock, or data lines at precisely timed moments during boot, they can slip custom code into memory before the operating system fully locks things down. The installation is delicate and very much non-trivial, but once in place, these chips can provide a semi-persistent custom firmware experience even on “patched” consoles.

More recently, devices like Mig Switch and Mig Flash took a different angle. They pose as ordinary game cards on the outside, but inside hold a microcontroller and a microSD slot. Using a separate dumper, a user can extract game images and associated metadata from legitimate cartridges onto the card. Pressing a button on the Mig cycles through those stored identities, presenting the console with a different “real” game every time. From the console’s perspective, it is just seeing one card being swapped for another; from the user’s perspective, an entire physical library collapses into a single cartridge.

Legally and ethically, this design walks straight into a minefield. It makes no attempt to distinguish between personal backups and downloaded images; it undermines the one-cartridge-one-license model that Nintendo relies on. Unsurprisingly, Nintendo’s reactions have involved both lawsuits against sellers and increasingly aggressive console-side detection and ban mechanisms, particularly as owners experiment with these devices on new-generation hardware.

Taken together, the Switch’s hack history reads like a case study in modern console control. Boot ROM bugs, joy-con rail shorting, microcontroller glitchers, identity-morphing flashcarts — each step in the chain reflects a tension between two conceptions of ownership. One side sees a purchased console as an object to be fully explored, modified, and repurposed; the other treats it as a managed endpoint in a larger service ecosystem, tightly bound by firmware updates and online checks.

Understanding the Switch After the Switch

With its successor now in the market, it is easier to see the original Switch in full. The new machine addresses many of the obvious complaints: higher performance, better screen options, more storage, stronger online infrastructure, a cleaner approach to backward compatibility. It is, in many ways, what people imagined a “next-gen hybrid” should be.

That makes the first Switch look even stranger and more interesting. It is not a technical marvel by contemporary standards. Joy-Con drift remains a practical annoyance. The OS feels frozen in time, minimalist to a fault, and conspicuously underpowered in online features. Internal storage is cramped, download speeds are unremarkable, and basic tasks like voice chat depend on awkward workarounds.

And yet, measured in cultural weight per watt, the Switch is hard to beat. It took a bet on a hybrid form factor when the rest of the industry was either doubling down on set-top-box designs or quietly retreating from dedicated hardware. It tried, in good faith and not always successfully, to fold motion control, haptics, cameras, and even cardboard into something more ambitious than a feature checklist. It survived long enough to become more than “this generation’s Nintendo console”: it became the background hardware of an entire era.

Looking back, the Switch feels less like one cleanly bounded product and more like a hinge: between disc and download, between living room and commute, between local control and cloud-tethered ownership. It stretched the idea of a console generation until the boundaries blurred, and in doing so, made the question “what is a console?” harder to answer in a way that fits neatly on a spec sheet.

Dual Video Outputs — A Brief Moment of Excess

On the back of the DOL-001 sit two video connectors:
one Analog AV Out, delivering composite and S-Video, and one Digital AV Out, capable of progressive-scan component or pure digital output.

This dual-port configuration allowed the GameCube to deliver some of the cleanest 480p video of its generation — when paired with Nintendo’s official component cable or modern GCVideo HDMI adapters.
Later revisions, beginning with the DOL-101, quietly removed the Digital AV Out entirely to cut cost, a change that would permanently limit video fidelity for subsequent models.

The Underside Map — Three Ports, One Vision

Flip the console over, and a second layer of design ambition appears.
Early GameCube units feature three expansion ports neatly hidden behind plastic covers:

1. High-Speed Port
2. Serial Port 1
3. Serial Port 2

While most players never used them, these interfaces were conceived to extend the GameCube far beyond its compact frame — a modular philosophy that anticipated modern “dock” ecosystems decades ahead of its time.

High-Speed Port — Gateway to the Game Boy Player

The High-Speed Port provided a 32 MHz parallel connection directly to the system’s internal bus.
Its most famous accessory, the Game Boy Player (DOL-017), attached underneath the console and effectively embedded a full Game Boy Advance system inside.

Unlike software emulation, the Player’s hardware ran native cartridges from the GB, GBC, and GBA generations with near-perfect compatibility.

However, the device required a boot disc to launch — one that was region-locked even though the hardware itself was not.
This limitation would later motivate the creation of community alternatives such as Game Boy Interface (GBI), which offered sharper video modes and lower input latency when loaded through homebrew launchers.

Serial Port 1 — From Dial-Up to Broadband

Serial Port 1 hosted the GameCube’s official networking peripherals:

• Modem Adapter (DOL-012) — a 56 K dial-up modem.
• Broadband Adapter (DOL-015) — a 10/100 Mbps Ethernet module.

These accessories enabled a handful of online and LAN titles such as Phantasy Star Online and Mario Kart: Double Dash!!.
While Nintendo’s online servers are long gone, the hardware remains functional — still used today for private LAN play and debugging.

Serial Port 2 — Forgotten Potential, Rediscovered Purpose

The second serial port was never used by any retail accessory.
Officially, it existed for potential debugging hardware, and on later revisions Nintendo simply deleted it altogether.
But for enthusiasts, that empty slot became the foundation of something far greater.

Modern homebrew developers repurposed the port through a tiny adapter known as SD2SP2 — a passive board that connects a microSD card directly to the console’s internal bus.
Inserted into Serial Port 2, it allows the GameCube to boot games and software directly from SD storage without modifying the hardware, turning a once-unused connector into a silent, high-speed drive bay.

Swiss and the “Disc-Free” Renaissance

To make use of SD2SP2, the GameCube must first load Swiss, the open-source Swiss-Army-knife of GameCube software.
Swiss acts as a universal loader and diagnostics toolkit, capable of launching ISOs, homebrew, and alternative boot programs such as GBI.

Launching Swiss originally required modchips or optical drive emulators, but the community surprisingly discovered a purely software-based exploit: using specific retail game discs and a modified memory card to execute code from saved data.

Among these exploitable titles is Super Smash Bros. Melee.

A crafted save file can overflow into executable memory, triggering Swiss without any soldering or hardware modification.
Other compatible games — including The Legend of Zelda: The Wind Waker, Twilight Princess, and Paper Mario: The Thousand-Year Door — offer similar pathways.

This method, often called a game-save exploit, transformed the GameCube into an open development platform using only official hardware: a retail disc, a memory card, and a microSD adapter.

Game Boy Player + GBI — The Refined Experience

Once Swiss is running, the Game Boy Interface replaces Nintendo’s original boot disc with a sharper, more responsive presentation.
Paired with the digital video output of the DOL-001, the result is arguably the highest-quality way to play Game Boy titles on original hardware — clean pixel scaling, optional 240p output, and near-zero lag.

What began as an obscure accessory has thus evolved into a professional-grade capture and archival platform, sustained by enthusiasts long after Nintendo discontinued support.

Legacy of an Expandable Machine

The early GameCube’s design reveals a philosophy Nintendo rarely revisited: a console built for future possibility.
Its dual video outputs and trio of bottom-mounted interfaces made it one of the most expandable mainstream consoles ever released.
Although most of those features went unused during its retail lifespan, they laid the groundwork for a vibrant modern community that continues to reinterpret the system through open-source tools and precision hardware.

Two decades later, the DOL-001 has outgrown the purpose Nintendo originally gave it. The ports that once seemed to lead nowhere now sustain an afterlife built by curiosity, engineering, and care.

References

1. Rodrigo Copetti, “Nintendo GameCube Architecture – A Practical Analysis,” Copetti.org, 2019. [Online]. Available: https://www.copetti.org/writings/consoles/gamecube/
2. “Booting Homebrew,” GC-Forever Wiki. [Online]. Available: https://www.gc-forever.com/wiki/index.php?title=Booting_homebrew
3. GameCube Homebrew – Game Save Exploits, GitHub repository, GameCubeHomebrew/GameSave-Exploits. [Online]. Available: https://github.com/GameCubeHomebrew/GameSave-Exploits
4. “SD2SP2 MicroSD Adapter for GameCube Serial Port 2,” StoneAgeGamer Store Listing. [Online]. Available: https://stoneagegamer.com/sd2sp2-micro-sd-card-adapter-1-2a-sdloader-for-gamecube-serial-port-2.html

My own unit is the white model, which only reinforces the “design object” feeling: it looks less like a toy and more like something that accidentally wandered out of a museum shop. It is a handheld that looks affordable, but behaves, and is priced, like a boutique instrument.

Two FPGAs Instead of One Emulator

Analogue’s core claim is simple: the Pocket is built around FPGAs, not a general-purpose CPU running software emulators.

A software emulator does something conceptually straightforward but technically messy: you take the original console’s behavior and model it in code. The emulator runs on a fast modern CPU, interpreting or recompiling instructions from the original system, faking video and audio timing, and approximating all the little quirks that games rely on. The host CPU is doing everything in sequence, one instruction after another, and the emulator is constantly juggling between accuracy and speed.

An FPGA works differently. It is a large grid of logic blocks and interconnects that can be configured to behave like custom hardware. Instead of “running an emulator”, you use a hardware description language (Verilog, VHDL, and so on) to describe the original console’s circuits: CPU, video hardware, audio unit, memory buses, timing generators. When the FPGA is configured with that design, it stops behaving like a generic chip and starts behaving like a hardware clone of the target console.

On the Pocket, that means:

• When a Game Boy core is loaded, the FPGA becomes a Game Boy-like machine at the electrical and timing level.
• When a Game Boy Advance core is loaded, the fabric is reconfigured into a different machine.
• When an openFPGA core targets a completely different system, the same physical silicon is rewired again into a new configuration.

The advantages over classic software emulation are subtle but important:

1. Timing and edge cases
   Original games often rely on undocumented hardware behavior, exact clock relationships, or tiny race conditions between CPU, video and audio. A traditional emulator has to model all of this in software, stepping through time in small increments and constantly approximating how several components would have behaved in parallel. An FPGA core, by contrast, is a network of parallel circuits. Clocks, phases, and signal propagations are part of the design itself, which makes strange timing tricks and borderline cases more likely to behave as they did on original hardware.

2. Latency and determinism
   A software emulator is stacked on top of an operating system, drivers, and a renderer. Frames may be delayed by buffering, scheduling, or synchronization with the host’s GPU. An FPGA system can be much more direct: input, game logic, and pixel output are chained together in hardware, with fewer layers in between. That tends to reduce input lag and produce more predictable frame pacing, especially when the output display supports the right refresh rates.

3. Hardware-level compatibility with cartridges and peripherals
   Because the FPGA recreates the original memory maps and bus protocols, plugging an original cartridge into the Pocket looks almost exactly like plugging it into the original console—from the game’s point of view. Save chips, mappers, and oddball enhancements are all accessed through the same kinds of signals they expect. This is much harder to reproduce cleanly when the only thing the emulator sees is a raw ROM file.

4. Separation between cores and platform
   In the FPGA model, each “core” is a hardware design. Improve the core and you ship a new bitstream; the Pocket simply loads it and becomes a slightly better version of the same console. Bugs and inaccuracies exist, but they take the form of fixable hardware descriptions rather than layers of clever hacks in C++ abstractions.

None of this is free. FPGA capacity is finite, development is slow, and the chips themselves are expensive. That cost is built into the Pocket’s price, and into the fact that it targets a relatively small audience who care about this level of fidelity.

The Screen: Ten Game Boys Worth of Resolution

The first time you see a Pocket in person, the screen dominates everything else. It is a 3.5-inch LCD running at 1600×1440, for an almost absurd 615 ppi in roughly the same footprint as the original Game Boy’s panel.

That resolution is not just a spec-sheet trophy. Classic handheld resolutions land neatly into integer scales: 160×144 from the DMG Game Boy becomes a 10×10 pixel grid per original pixel. Sprites and text suddenly look like clean high-resolution artwork instead of shimmering blocks. Analogue leans into that with a library of display modes: DMG-style greens, Game Boy Color palettes, GBA-style warmer tones, different sub-pixel patterns, simulated grids and subtle “LCD” behaviors for those who miss the original flaws.

Combined with precise timing on the FPGA side, motion feels crisp and immediate. There is no sense of the smeared ghosting that haunts many cheaper LCD mods; instead, you get a kind of hyper-real version of handheld history—more detailed than the original screens ever were, but still visually anchored to their character.

Industrial Design: A Minimalist Tribute, Not a Clone

From a distance, the Pocket reads as “Game Boy, but drawn with a ruler”. The body is a clean rectangle with a high-gloss glass front covering the entire upper half and gently rounded edges on the shell. Face buttons follow Nintendo’s basic layout but are tighter and more symmetrical, with Start and Select tucked into the lower bezel instead of floating somewhere in the middle.

On the sides and back, the practicalities appear: a link port, microSD slot, USB-C for charging and the dock, and a rear cartridge slot that accepts all Game Boy-family Game Pak types.

This device feels even more like a minimalist reinterpretation than a straightforward homage. It sits comfortably on a desk next to cameras, keyboards, and laptops without looking out of place. It is a handheld, but it is also a piece of industrial design that happens to play cartridges.

Official Cartridge Adapters: Beyond the Game Boy Family

Out of the box, the Pocket is a Game Boy, Game Boy Color, and Game Boy Advance machine. Officially, though, Analogue never meant it to stop there. The company sells a separate adapter ecosystem that lets the same FPGA brain impersonate a wider set of handhelds while still using original cartridges.

There are two main branches:

• A Game Gear adapter, sold as a dedicated accessory.
• The Analogue Pocket Adapter Set, a three-pack that covers Neo Geo Pocket Color, Atari Lynx, and TurboGrafx-16 / PC Engine / SuperGrafx HuCards.

Inserted into the Pocket’s rear slot, these adapters become little plastic bridges between very different ecosystems. In practical use, they feel surprisingly seamless. Slide in a Lynx card or Neo Geo Pocket Color cart and the system reconfigures to match the original hardware model, just as it does switching from GB to GBA.

Owning the Analogue Pocket with these official adapters turns it into a strangely dense archive: Atari’s early color experiments, SNK’s compact fighting machine, Hudson’s HuCard world—all squeezed into the footprint of a modern Game Boy. It is still “original media, original timing”, but now spread across several companies that never designed their hardware to coexist like this.

openFPGA: From Cartridges to ROM-Driven Multi-System

The physical cartridges are only half the story. The other half lives on the microSD card and inside the firmware.

Analogue’s openFPGA initiative exposes the Pocket’s FPGA fabric to third-party developers. Instead of only relying on Analogue’s own cores, openFPGA allows independent projects to define their own hardware implementations and ship them as downloadable cores.

Once configured, the Pocket is no longer just a Game Boy derivative with adapters. It becomes a menu of soft consoles and arcade boards:

• Additional handhelds beyond the official adapter set.
• Eight- and sixteen-bit home consoles like NES, SNES, Mega Drive / Genesis, and PC Engine in core form.
• Selected arcade systems, where FPGA resources permit.

ROMs live on the microSD card; the Pocket’s role becomes that of a hardware-accurate host. For preservation, this is an appealing model. Cores can evolve, bug-fix, and be archived independently of any given hardware revision, while ROM sets safeguard aging carts whose save batteries will eventually die.

The ethical line is familiar: nothing in the hardware stops you from loading whatever ROMs you can find. Analogue keeps its own OS and official cores closed, and frames openFPGA as a platform for development and preservation. What people actually load into it is, as always, their own responsibility.

The Dock: From Handheld to Living-Room FPGA Console

The Pocket experience changes character when you drop the handheld into the Analogue Dock. Physically, the dock is understated: a compact base with a USB-C post at the top, HDMI and USB-A ports at the back, and not much else to look at.

Once docked, the Pocket’s FPGA cores drive 1080p HDMI output to a TV or capture device. Wired USB controllers and a range of Bluetooth pads can be paired, turning handheld games into couch multiplayer. Firmware updates have gradually broadened controller support, including Nintendo’s own retro-style wireless pads intended for Switch Online, which fit naturally when you are running SNES or Mega Drive cores via ROMs.

For Game Boy-era software, the dock effectively recreates the old “handheld-on-TV” tricks—think Super Game Boy on the Super Famicom or Game Boy Player on the GameCube—but with cleaner video, more flexible scaling, and modern controller options. For openFPGA home console cores, the Pocket-in-a-dock simply behaves like a small FPGA console that happens to undock and become a handheld when you leave the house.

The economics are still harsh. A 219.99 USD handheld, a 99.99 USD dock, a 99.99 USD adapter set, a separate Game Gear adapter, and whatever controllers you bring: this is not a budget way to revisit the past. It is, very clearly, an enthusiast stack.

A Machine Honestly Aimed at Enthusiasts

There are cheaper, simpler ways to play ROMs. Android handhelds, Raspberry Pi builds, and a whole ecosystem of emulation-focused devices exist precisely to treat retro gaming as a solved software problem: pick a SoC, run Linux, ship an emulator front-end.

The Analogue Pocket, once completed with its official adapters and dock, is almost aggressively different. It uses FPGAs where a cheap ARM chip would suffice, a screen whose pixel density has no business being attached to thirty-year-old games, and a set of accessories that clearly assume you are prepared to spend serious money on a hobby that already peaked decades ago.

For some people, this will forever be the wrong answer. If all you want is a handful of childhood favorites on the commute, an inexpensive emulator handheld makes more practical sense. But if you treat cartridges as a collection, care about timing quirks, want Atari Lynx cards and HuCards and Neo Geo Pocket Color carts to all converge on one piece of hardware, and appreciate the sensation of over-specified design in your hands, the Pocket becomes a very particular kind of ideal object.

In that sense, it fits neatly into what Retro Obsession is about: not merely keeping old games technically playable, but asking what happens when modern engineering encounters old ideas without treating them as disposable. The Analogue Pocket is one of the clearest products of that encounter — an FPGA handheld that has grown far beyond the Game Boy family, unapologetically priced and designed for the sort of person who sees such a machine and immediately imagines the shelf it should live on.

Both are Japanese versions, brand new and unopened — intended as part of my permanent collection.

The Mario Edition

The Super Mario Bros. version includes four programs:

1. Super Mario Bros. (Famicom/NES version)
2. Super Mario Bros. 2 (the Japanese Famicom Disk System version, known as The Lost Levels overseas)
3. Ball — the original 1980 Game & Watch juggling game
4. A charming digital clock mode, featuring dynamic Mario animations

It’s an elegant little tribute that bridges the earliest Game & Watch era with the golden age of the Famicom.

The Zelda Edition

The Legend of Zelda version follows the same concept but with a richer library:

1. The Legend of Zelda (Famicom version)
2. Zelda II: The Adventure of Link
3. Link’s Awakening (Game Boy version, in its original monochrome form)
4. A special playable clock and timer, themed around Hyrule

It’s more than nostalgia — it’s a playable museum piece.

Hidden Potential

Interestingly, both models are capable of more than what’s officially included.
By following community-developed hardware modification guides, these systems can be customized to run additional titles.
Their compact form factor and ARM-based architecture make them a fascinating platform for enthusiasts to explore.

A Worthy Addition

Whether they are worth owning depends on how much value you place on an object that is part toy, part anniversary piece, and part invitation to tinker. For me, the more affordable Zelda edition makes the clearest case. Both machines look back to Nintendo’s beginnings without pretending to belong to them: modern objects carrying an old silhouette, with just enough technological curiosity beneath the surface.

Anyone who seriously plays on original hardware eventually runs into the same problem: classic consoles output stubborn analog video formats designed for 15 kHz CRTs, while modern displays speak almost exclusively HDMI. Composite video, RGB SCART, YPbPr component, and VGA all belong to another era; contemporary TVs and monitors either support them poorly or not at all.

One option is to chain generic adapters and let the TV guess what to do. Another is to buy a dedicated high-end scaler. GBSC (often written “GBS-C”, short for “GBS + Control”) takes a different path: it starts from a cheap GBS-8200 arcade converter board, replaces its control logic with an open-source firmware called gbs-control, and turns it into a low-lag, multi-format scaler that can feed clean VGA or HDMI to modern displays.

I now have one of the integrated GBSC all-in-one units in a blue acrylic shell. Under that acrylic, it is still a GBS-8200-class board at heart, but with the usual hardware modifications and the custom firmware already in place, ready to sit between a stack of retro consoles and a single modern screen.

From GBS-8200 to GBSC

The starting point for this whole ecosystem is the GBS-8200 / 8220 series: inexpensive Chinese arcade converter boards originally intended to adapt 15 kHz CGA/EGA or YUV signals to 31 kHz VGA monitors. Their job is to double the scan rate so that low-frequency arcade PCBs or home computers can drive a standard VGA input.

At the core of these boards sits the Tvia TrueView TV5725 scaler chip, a multimedia processor capable of handling a wide range of analog sources and output resolutions. In stock form, the firmware that drives the TV5725 is conservative: it does the bare minimum to convert 15 kHz RGB or YPbPr into a usable VGA image, but with visible artefacts, inconsistent deinterlacing, and input lag that is not ideal for responsive gameplay.

The gbs-control project intervenes at this exact point. It replaces the original control software with a new firmware that runs on an external microcontroller and directly reprograms the TV5725 for the specific task of retro gaming. The official documentation describes it as an “alternative firmware for Tvia TrueView5725 based upscalers / video converter boards” with features such as very low lag, sharp upscaling, and stable 240p/480i switching.

In the canonical build, gbs-control runs on an ESP8266 Wi-Fi microcontroller (for example a Wemos D1 or NodeMCU), mounted onto the GBS-8200 via a small adapter board. The ESP8266 takes over configuration of the scaler, exposes a web-based GUI, and implements all the new processing logic. Once this setup is in place, the resulting device is usually referred to as GBSC or GBS-C.

Inputs, Outputs, and What GBSC Actually Connects

Conceptually, GBSC is a bridge between legacy analog formats and a small set of clean, predictable outputs.

On the input side, the underlying GBS-8200 hardware accepts a mix of low-frequency signals: CGA/EGA-style RGBS at 15–25 kHz, VGA-style RGBHV at 31 kHz, and YPbPr component in common video modes such as 480i, 480p, 720p, and 1080i/1080p. Depending on the specific GBSC variant, these are presented as:

• a Euro-SCART or JP-21 RGBS connector for consoles and computers with RGB outputs;
• RCA-style jacks, wired for direct RGBS or arcade-style RGB inputs;
• YPbPr component jacks for devices such as PlayStation 2, original Xbox, and Wii;
• a VGA-style input for Dreamcast VGA boxes, PCs, or other RGBHV sources.

Some all-in-one designs add AV, composite or S-Video (such as my own unit) via extra analog front-end circuitry, but that is an implementation detail of individual products rather than a universal feature of the firmware. Other commercialized units, such as the ODV-GBS-C, explicitly limit themselves to RGBS SCART, component YPbPr, and VGA/RGBHV as their supported inputs.

On the output side, a GBSC build typically provides:

• an analog VGA or RGBHV output suitable for CRT VGA monitors, PVM/BVMs (with appropriate cabling), or further analog routing;
• a digital HDMI output, added via a companion converter board in many AIO/Pro units, with resolutions up to 1080p.

This makes it possible to route essentially an entire 8-/16-/32-bit console collection, plus some early 2000s systems, through a single scaler and into a modern TV or capture card.

Upscaling: From 240p to High Definition

The main workload for GBSC is still upscaling 240p and 480i sources to progressive, higher-resolution outputs.

With gbs-control in charge, 15 kHz “240p/288p” content from consoles and microcomputers can be line-doubled and scaled to common PC monitor formats such as 640×480, 800×600, 1024×768, or 1360×768, and then further adapted to 720p or 1080p for HDMI. The firmware exposes controls for scaling, cropping, and timing in a web GUI, making it straightforward to dial in a profile where pixels are mapped cleanly and aspect ratios are correct.

A key feature compared to the stock GBS-8200 firmware is very low lag and robust handling of mixed 240p/480i content. The official feature list for gbs-control emphasizes that output timing is decoupled from input timing, so 240p↔480i transitions (for example, when a game switches between an interlaced menu and a progressive in-game image) do not cause the output to lose sync. For systems such as PlayStation or Saturn, where this pattern is common, this is a practical improvement over many generic HDMI converters.

Motion-adaptive deinterlacing is applied “on demand” to 480i sources, reducing combing artefacts on moving edges without excessively blurring static parts of the picture.Combined with adjustable sharpness and gamma, this allows GBSC to produce a crisp but not excessively harsh image that many users consider comparable to more expensive scalers when configured correctly.

In practice, a typical configuration might map 240p game consoles to 960p with scanline emulation on, PS2 component output to 720p with motion-adaptive deinterlacing, and original Xbox to 1080p, all with per-input profiles stored on the device.

Downscaling and 240p Experiments

An interesting capability of gbs-control, and one that pushes the TV5725 beyond its original design brief, is its support for downscaling higher-resolution signals back to 240p-class outputs.

Used this way, GBSC can take 480p or higher signals from later consoles or even PCs and generate a 240p-style output suitable for CRT televisions and monitors. Documentation and user reports describe use cases such as feeding 240p from GBSC into another scaler, or deliberately downscaling HD re-releases to 240p to recover something closer to the look of their original low-resolution counterparts.

This is a niche but technically notable feature: the same scaler chip that line-doubles 240p for LCDs can also be configured to act as a kind of “reverse scan converter”, letting modern hardware drive legacy CRTs at genuinely low scan frequencies.

DIY Builds vs. All-In-One Units

The original gbs-control documentation assumes a DIY build: a bare GBS-8200 board, an ESP8266 module, a few supporting components (such as a Si5351A clock generator and power-supply clean-up parts), plus wiring and optional extras like an OLED status display. The wiki includes detailed notes on hardware variations, analog front-end modifications, clock generator installation, and other improvements that bring the board’s analogue performance closer to what the firmware can deliver.

Over time, community and commercial designs have converged toward more integrated solutions. Projects such as GBS-C AIO define an add-on board that sits on top of the GBS-8200, consolidating the ESP8266, clock generator, input connectors, power regulation, and often an HDMI converter into a single neat assembly. Commercial devices like the ODV-GBS-C and GBSC Pro take this one step further by shipping a complete, enclosed scaler with gbs-control preinstalled, SCART/YPbPr/VGA inputs, VGA/HDMI outputs, and infrared remote control, positioned explicitly as low-lag upscalers for retro game consoles.

My own blue acrylic GBSC belongs to this integrated family: the original GBS-class board is still there, but all of the common hardware modifications and the ESP8266 controller are already part of a single stack, which means setup is essentially reduced to cabling and web configuration rather than soldering.

In that sense, GBSC as a platform covers a spectrum. At one end, it is a highly hackable combination of off-the-shelf PCB and open firmware, inviting hardware and software experimentation. At the other, it is moving toward a semi-standardized “retro scaler box” product layer, where the open firmware quietly drives a device sold to users who may never touch a soldering iron.

Performance, Limitations, and Niche

Compared to generic “AV-to-HDMI” adapters, a well-built GBSC offers much lower input lag, more faithful handling of 240p and 480i, and predictable timing. Less than one frame of delay is typical for upscaling chains based on gbs-control and GBS-C AIO-class hardware, which is sufficient for fast action games and shooters. Compatibility with mainstream 8-bit, 16-bit, 32-bit, and early 2000s consoles and microcomputers is strong, especially over RGB and YPbPr.

At the same time, GBSC does not completely replace higher-end commercial scalers. Some exotic video modes and marginal signal conditions remain challenging. Noise and stability still depend on the quality of the underlying GBS-8200 analogue front-end and power supply, which can vary between board revisions and manufacturers. Downscaling features, while powerful, require careful configuration and may not behave identically across all sources.

From an ecosystem perspective, GBSC fills a clear niche:

• cheaper and more flexible than most dedicated, proprietary scalers;
• significantly more capable than generic adapters;
• and open enough that the community can continue to refine both hardware and firmware over time.

GBSC occupies the narrow span between analog and HDMI, and between ad-hoc DIY tinkering and the increasingly formal world of retro-video products.

Reference

GBS-Control Repository

This is not a neutral console. The N64 takes sides—in storage, in input, in memory—and lives with the consequences.

A 3D Machine Built Around Memory and Microcode

Under the plastic shell, the N64 is not a simple “faster SNES”. It is closer to a tiny workstation condensed for the living room.

A 64-bit NEC VR4300 CPU sits next to the Reality Coprocessor, which houses two cores of its own: the Reality Signal Processor, running vector microcode, and the Reality Display Processor, responsible for rasterization, blending, depth buffering, and a good deal of the system’s characteristic “soft” look. Games do not just call a fixed set of GPU functions; they push display lists and, in some cases, custom microcode that literally rewires how the hardware pipeline behaves.

That flexibility is intoxicating and expensive. It made effects like smooth Z-buffered 3D, complex particle work, and aggressive anti-aliasing possible on mid-90s hardware. It also made development harder, and it is precisely why, decades later, N64 emulation can still feel slightly off—one game looks perfect, the next has subtly wrong fog, broken blending, or timing that just doesn’t match real hardware.

The N64 is a console that never fully fits inside an emulator window. There is always some microcode, some timing edge case, slightly out of reach.

Expansion Pak and the Strange, Layered Memory Story

Nintendo did not just throw RAM at the problem and call it a day. The N64’s memory story is layered and oddly personal.

Base units ship with 4 MB of RDRAM and a “Jumper Pak” plugged into a slot on the top. That module is essentially a terminator so the memory bus behaves properly. The real magic happens when you pull it out and slide in the bright red Expansion Pak, doubling main memory to 8 MB. Suddenly, certain games gain sharper modes, more complex geometry, or entire features that are simply unavailable on the base machine. Some late-generation titles quietly flip the relationship entirely and refuse to start without the upgrade.

Originally, that extra RAM was tied to the 64DD—the magnetic disk add-on that was supposed to solve the cartridge’s capacity problem and turn the N64 into something closer to a writable, persistent world machine. The 64DD came late, stayed in Japan, and left a handful of experimental titles plus a strange alternate timeline in which the N64 might have had a second life as a disk-based console. The Expansion Pak survived and became a kind of mid-generation spec bump, a small apology from the hardware to the software.

Memory on N64 is not just soldered to the board. It appears in the console, in the expansion slot, in the cartridge as save chips, and in the controller as a memory card. That distribution matters, because it influences how games think about data: what lives permanently in a cart, what moves with the player, what is considered expendable.

The Controller as a Tiny Expansion Bus

Most people remember the N64 controller for its shape. But if you flip it over, the most interesting part is the little slot on the back.

That slot is a miniature expansion bus. Into it go the Controller Pak (memory card), Rumble Pak (force feedback), Transfer Pak (Game Boy connectivity), and a zoo of third-party experiments. Nintendo effectively moved part of the system’s state into the player’s hands: saves, rumble, cross-platform data all live in that cavity under your fingers.

The Controller Pak is where the N64 flirts with the memory card culture that disc-based consoles made standard. Official cards store only tens of kilobytes, divided into pages, with a tiny save manager you can access from some games or from the console’s own UI. On paper, it is elegant: you can simply unplug your controller and walk to a friend’s house, taking your ghosts, teams, and settings with you.

In practice, the split is messy. Many first-party titles use save chips on the cartridge itself and ignore the Controller Pak entirely. Others lean heavily on the Pak, filling it quickly and forcing players to micromanage blocks in an interface that feels more like an engineering tool than a consumer product. It is the kind of design you get when hardware engineers imagine how people will share data in physical space—and nearly get it right.

::: details, title="N64 Controller Port and Joybus Frame Format"

Behind that friendly row of four controller ports on the front of the Nintendo 64 is an extremely minimalist design: each port only exposes three lines—power, data, and ground. Power is 3.3 V (not the 5 V people often expect), ground is shared, and everything else has to squeeze through that single data line: button states, analog stick position, even save data to and from the Controller Pak all ride on the same wire.

Nintendo built a custom protocol for this called Joybus: a single-wire, half-duplex, open-drain serial bus. When idle, the line is held high at 3.3 V by a pull-up resistor. When either the console or the controller wants to talk, it only ever drives the line low, then releases it and lets the pull-up bring the voltage back up. Logical 0 and 1 are not distinguished by voltage levels but by pulse width: each bit occupies a 4 µs time slot, with a longer low pulse meaning one value and a shorter low pulse meaning the other. It looks deceptively simple on the outside, but internally it demands very strict timing.

The console is always the master on this bus. To read the controller, it first sends a short Joybus command on the data line—most commonly 0x01, “give me your current state”—then releases the line and switches to receive mode. The controller replies with a 32-bit block: the first two bytes encode button states, the next two bytes encode the analog stick’s X/Y coordinates as signed values around a calibrated center. Other commands go deeper: there are opcodes to read and write 32-byte blocks to whatever is plugged into the expansion slot (Controller Pak, Rumble Pak, etc.), but from the console’s point of view it is always just talking over the same single wire.

The expansion slot on the underside of the controller adds another layer of indirection. Internally, that slot is wired as a small parallel bus—with address lines, data lines, and read/write control signals—much closer to a traditional memory interface. The controller’s microcontroller acts as a bridge: when the console issues Joybus commands to read or write the Controller Pak or Rumble Pak, the controller translates those serialized requests into parallel accesses to the expansion device, then wraps the results back into Joybus frames for the console. In other words, the N64 never talks directly to the memory card or the rumble motor; it only talks to the controller, and the controller quietly runs a tiny storage and I/O network on its behalf.

The entire exchange is made up of 4 µs bit time slots: long-low/short-high encodes a 0, short-low/long-high encodes a 1. All bytes and the stop bit follow this rule.

Poll Command Frame Format

Console → Controller

Controller State Response Frame

Controller → Console

Button Bit Mapping (First Two Bytes of 0x01 Response)

Note: in the N64 protocol, pressed = 1, not pressed = 0.

:::

Rumble Pak: When the Controller Started Talking Back

Force feedback existed before the N64, but the Rumble Pak is where it became part of the mainstream vocabulary of console hardware.

Bundled with Star Fox 64, the Rumble Pak rides in the same slot as the Controller Pak and turns impact into a physical event. Missiles, collisions, explosions—they all get translated into that now-iconic buzzing weight in your palms. Suddenly the controller is no longer a mute intermediary; it has opinions about what just happened on screen.

The design choice is very Nintendo 64: bold but awkward. Because Rumble Pak and Controller Pak share the same port, many games make you choose: do you want your saves or your rumble? Some titles build in elaborate hot-swap screens just so you can switch between memory and feedback at the right time.

Third-party manufacturers eventually responded with hybrid units—rumble plus memory in one shell, more pages per card, different battery schemes—but the tension never quite goes away. The console is pulling in two directions at once: toward a controller that stores your identity, and toward one that physically signals what the game feels like.

Analog Stick: The Breakthrough and the Flaw

At the center of it all sits the analog stick. If the SNES defined what a modern digital pad looks like, the N64 defined what 3D movement feels like when you first put down the D-pad.

The stick is placed on the middle prong, with the Z-trigger just behind it. It is meant to be the default hand position for 3D: you wrap your left hand around the analog grip, right hand on the buttons, and suddenly Super Mario 64 makes sense. You are not pressing “right” anymore; you are leaning into a direction, modulating speed and angle, exploring a world instead of sliding along an invisible grid.

The stick’s impact on design is hard to overstate. It makes subtle differences in movement—tiptoe, walk, run—feel natural. It works in tandem with camera systems that are still learning how to exist. It lays the conceptual foundation for everything from 3D platformers to console first-person shooters.

Mechanically, though, it is fragile. The original N64 analog module uses optical encoding and a plastic cage that wears itself down as you play. Gates get ground into ovals, powdered plastic accumulates inside, and the neutral position drifts over time. Anyone who grew up with Mario Party remembers not just the minigames, but the way they ate controllers for breakfast.

There is also an architectural limitation baked into the design: one stick, not two. Camera control is mapped to digital C-buttons, forcing developers into clever workarounds—lock-on systems, snap turning, context-sensitive cameras that try to read your intent from limited input. In some ways this constraint forced some of the most elegant 3D control schemes of the era; in others it makes the N64 feel like a prototype for the dual-stick future that would arrive with the next generation.

Today, replacement stick modules are a whole subculture. Some chase the original feel with better materials, others swap in modern gimbals while preserving the console’s expectations. Very few N64 controllers in circulation still feel like they did in 1996.

A Layout That Taught People How to Live in 3D

If you map the N64 controller on paper, it looks wrong. In motion, it quietly retrains your hands.

Super Mario 64 teaches you that the analog stick is your primary verb and that the camera is a somewhat stubborn partner you negotiate with. The Legend of Zelda: Ocarina of Time introduces Z-targeting and turns lock-on into a way of thinking about 3D combat: you are not just swinging in space, you are focusing on a relationship between two bodies in a volume. GoldenEye 007 and Perfect Dark try out different combinations of stick, C-buttons, and shoulder inputs, searching for a console-friendly FPS vocabulary.

None of these control schemes are quite “final”. But together they define a language: analog as default movement, triggers as primary actions, face buttons as context, camera as something that can be partially automated. The exact layouts change; the mental model stays.

Add in the four controller ports on the front of the console and the effect multiplies. N64 multiplayer is not just about four rectangles of split screen; it is about four people learning 3D space together, with rumble and analog and memory cards all in the loop. It is hard, even now, to replicate the specific feel of four N64 controllers tangled around a CRT.

Cartridges, 64DD, and the Cost of Saying “No” to CDs

All of this sits on top of a storage decision that time has not been kind to.

Where Sony and Sega embraced cheap CD-ROMs, Nintendo chose to stay with proprietary cartridges. The argument was straightforward: cartridges are fast, robust, and difficult to pirate. Load times vanish. Games feel immediate. If you care about streaming 3D worlds without obvious breaks, that is a tempting trade.

The trade, however, comes with a bill:

• Cartridges are small. In an era when CDs offer hundreds of megabytes, an N64 cart might give you a handful of tens.
• They are expensive to manufacture and slow to produce. Misjudge demand and you either leave money on the table or sit on unsold ROM.
• They shift risk onto publishers, who must commit capital earlier and in larger chunks.
• They make FMV, rich voice acting, and lavish audio a luxury rather than a baseline expectation.

The 64DD is the hardware manifestation of a lingering doubt. Somewhere inside Nintendo there was a recognition that cartridges alone could not carry the platform forever. The disk drive bolts under the console, brings rewritable 64 MB disks, a real-time clock, and the idea of online connectivity. It is a fascinating device in isolation and a quiet admission that the original media decision had painted the N64 into a corner.

The 64DD’s limited release and tiny library turn it into a historical footnote. But it highlights the central tension: the N64’s cartridge strategy gave it speed and solidity at the cost of breadth. Its library is a sharp, narrow mountain rather than a sprawling landscape. The games that made it are extraordinary. The ones that never arrived tell an equally important story.

Why Emulation Never Feels Entirely Finished

For many other consoles, “good enough” emulation arrived and quietly solidified. The N64 has been stubborn.

To emulate it faithfully, you must handle not just a MIPS CPU and some standard peripherals, but the way that CPU, the RSP, and the RDP all talk to each other—with microcode sitting in between as negotiator and troublemaker. Many emulators historically took a high-level approach: detect which microcode variant a game uses and map its calls to hand-written, high-level implementations. This works beautifully for popular, well-understood games and falls apart the moment something unusual happens.

More recent low-level efforts get closer to the metal, but they bring the cost of complexity and performance. You are trying to recreate a mid-90s experiment in flexible graphics hardware whose official documentation never fully told the whole truth.

The result, even now, is that “N64 emulation” is a spectrum. Some titles are indistinguishable from real hardware to all but the most obsessive eyes; others remain stubborn in small, uncanny ways. It is not that the N64 refuses to be emulated; it is that its personality is encoded in the cracks between components.

Flash Carts, Mods, and Living With an N64 in 2024

If you actually want to use a Nintendo 64 today rather than just own one, the machine you interact with is almost always a hybrid.

Flash cartridges like EverDrive 64 and 64Drive turn the cart slot into a general-purpose loader. An SD card quietly replaces shelves of aging plastic; save data moves into files; fan translations, romhacks, and homebrew sit next to first-party classics. The original cartridge format, once a bottleneck, becomes a shell wrapped around something remarkably flexible.

Controller Paks get swapped out for modern, battery-free designs. Analog sticks are replaced, sometimes with new modules that preserve the original profile, sometimes with entirely different mechanisms. The Expansion Pak becomes permanent; the Jumper Pak gets retired to a drawer as a relic of a four-megabyte world.

Video outputs are modded—RGB, component, HDMI—and routed through scalers that try to square the circle between a console built for CRT phosphors and a living room full of LCD panels. On the other end, FPGA re-implementations and high-accuracy emulators provide a parallel line of preservation, offering “N64-like” timing and behavior for those who cannot or do not want to maintain original hardware.

The N64 that actually gets played in 2025 is rarely stock. It is a dialogue between 1996’s idea of the future and our current expectations of convenience.

A High-Risk Machine Worth Remembering

If you judge the Nintendo 64 purely by market share, it is the generation where Nintendo stopped being the default home console. The cartridge strategy, the 64DD detour, the constrained third-party library—all of that reads like a case study in how to lose ground.

But if you trace the lines of influence, the picture changes.
The N64’s analog stick and trigger layout quietly defined how 3D movement should feel. Its controller slot turned peripherals into part of everyday play, from memory cards to rumble. Its layered memory design, from Expansion Pak to Controller Pak to on-cart saves, hinted at a world where hardware specs and player identity could both evolve mid-generation. Its microcode-driven architecture gave us a console that is still teaching emulator authors humility.

It is not a comfortable machine, and that is exactly why it is interesting.
The Nintendo 64 is what happens when a company absolutely believes in 3D and is willing to take real risks on how to get there, even when those risks do not pay off cleanly. For better and worse, it is high-risk hardware in the move to 3D—and that is precisely what makes it worth revisiting today.

Finding original game cartridges or discs has become increasingly difficult.
Even if you manage to find the title you want, there’s no guarantee it will work.
Most of these games are in Japanese or English, which poses an additional challenge for players from other regions. As a result, many enthusiasts resort to flash cartridges or burnable discs just to experience them.

Lost Instructions in a Connected Age

While many communities and organizations have archived classic game ROMs, most of these dumps are designed for emulators, not for use on the original hardware.
The real problem lies in the knowledge gap:
finding the necessary guides, tools, and software for pre-internet-era devices is getting harder every year.

Some of the once-popular websites are now gone;
tutorials are fragmented or outdated;
and the key utilities they reference have changed beyond recognition.

Even when everything seems to be available, hardware from that era was rarely standardized or user-friendly — meaning that, even with all the information in hand, achieving your goal can still be surprisingly difficult.

The Skill That’s Disappearing

I’ve archived ROMs for most mainstream consoles, and most of them still run fine on emulators.
But that doesn’t solve the core problem — how to play them on the original machines.
In fact, figuring out how to make old consoles work has quietly become a specialized skill, one that fewer and fewer people still practice.

Emulators have enjoyed a huge and growing audience, but the number of people who actually play on physical hardware continues to shrink.
Part of that is due to aging components and failing drives, but there’s also a social factor: older players move on, and newer ones often give up quickly, leaving behind an entire generation of hardware whose operating knowledge fades into obscurity.

When “Just Use an Emulator” Isn’t Enough

Take the PlayStation 1, for instance.
Everyone knows it can be modded, but there are few comprehensive resources comparing the different mod methods — their pros, cons, and compatibility issues.
Systematic, updated tutorials are even rarer.

And this isn’t just a problem for consoles from the 1990s.
Consider the Nintendo 3DS, a handheld from the previous generation.
During its lifetime, there were two major hacking approaches:

• Hardware modification using flashcards like Gateway or Sky3DS
• Software modification through the Boot9Strap (b9s) exploit

As b9s became mainstream, flashcard-based methods fell into obscurity.
Today, many players dismiss them as “unnecessary.”

When I try to research how to softmod a console, burn discs for PS1, or revive old flashcards for 3DS, the most common responses I get are:

“Just solder a chip, it’s cheaper.”
“Why go through all that trouble?”
“Just use an emulator — it’s the same thing.”

But is it really the same?

The Difference Between Playing and Understanding

For many systems before the PS2 era, emulators can reproduce a remarkable share of the original experience. They’re accessible, reliable, and in many ways superior to the real thing.

But that’s not my goal.
I am less interested in merely being able to play these games than in the opportunity to study the history of console engineering through firsthand experience. To feel the quirks of the original hardware, to understand the design logic that shaped an entire generation.

That’s what’s being lost.
And as tools disappear, communities fade, and collective memory erodes, that loss feels deeper than simple nostalgia.

Sometimes, I can’t help but feel powerless.
Maybe compromise is inevitable.
But I still hope that somewhere, someone will keep the knowledge alive —
so that “how to play” never becomes a forgotten language.

Industrial Design and Aesthetic Direction

The Micro is one of the rare moments when Nintendo prioritized design sophistication over mass-market pragmatism. Its anodized aluminum front panel, its glossy and interchangeable faceplates, and its starkly minimal rectangular frame projected an aesthetic closer to early-2000s portable media players than to Nintendo’s toy-like plastic heritage. The device feels precise, metallic, and unusually mature for a Game Boy.

This elegance, however, came with a price. The mirror-finished faceplates scratch easily, and the metallic surfaces can chip or lose their original sheen with regular use. Units in pristine condition have therefore become increasingly scarce, contributing to rising market value. My own unit—well-preserved, complete-in-box, and free from major cosmetic wear—cost approximately 400 USD. Its condition and completeness significantly contributed to the price, reflecting the growing scarcity of pristine Micro systems.

Display Quality: The Sharpest GBA Screen

Despite its small size, the Game Boy Micro features what many consider the finest display ever put into a GBA-family system. Its high pixel density compresses the 240×160 resolution into a remarkably sharp and vivid presentation. Sprites gain a refined crispness, background tiles appear cleanly delineated, and the characteristic dithering used in GBA graphics blends more naturally into the scene. Colors are vivid yet controlled, avoiding the artificial saturation seen in some modern IPS retrofits.

This produces an experience that feels unexpectedly modern—almost as if GBA games were being displayed on a contemporary miniature screen. The limitation, of course, lies in the physical size: text-heavy RPGs or games that rely on detailed interfaces can become challenging to read, and extended play may strain the eyes. The Micro excels not in scale but in precision.

Handfeel and Ergonomics

The Micro’s form factor suggests discomfort at first glance, but Nintendo’s attention to tactile detail becomes evident quickly. The D-pad has a short, precise pivot that feels distinctly mechanical, and the A/B buttons offer crisp feedback uncommon even among other Nintendo handhelds. The device’s weight makes it pleasant to hold for shorter sessions, and the responsiveness of the controls reinforces its premium feel.

However, the constraints imposed by its size cannot be ignored. Players with larger hands may experience fatigue, and long sessions with RPGs or platformers can become uncomfortable. The Micro is best suited to fast, immediate gameplay rather than multi-hour commitments. It is a device optimized for short, satisfying bursts of interaction rather than sustained engagements.

Hardware Architecture and the Absence of Backward Compatibility

The Micro’s inability to play Game Boy and Game Boy Color cartridges is often misinterpreted as a limitation, yet it is actually the result of a strategic engineering choice. Historically, Nintendo provided backward compatibility through the physical inclusion of earlier hardware. The Game Boy Color retained a Z80-compatible processor to play original GB titles, and the Game Boy Advance incorporated much of the GBC’s graphics and timing circuitry, allowing it to operate in a true GBC mode.

This approach offered excellent compatibility but required substantial board space, deeper cartridge slots, and additional electrical pathways. The Micro was built around the opposite philosophy. To achieve its unusually compact size, Nintendo removed all GB/GBC-related circuitry, leaving only the ARM7 architecture and the necessary components for native GBA operation. Without the older hardware blocks, compatibility is not simply disabled—it becomes electrically impossible. The Micro’s form factor is inseparable from this decision.

Market Reception and Collector Re-evaluation

At launch, the Game Boy Micro struggled. Its high price, small display, fragile finish, and unclear place in the market left consumers confused, especially as the Nintendo DS surged in popularity. Many players could not justify a premium GBA variant that sacrificed backward compatibility and practicality.

Yet as time passed, the Micro’s unique qualities transformed it from a commercial oddity into a collector’s treasure. Its scarcity, refined industrial design, precision display, and vulnerability to cosmetic damage have all contributed to its rising cultural and monetary value. Special editions like the Famicom and Mother 3 variants, in particular, have become highly prized artifacts within the retro hardware community.

Historical Significance

Viewed in retrospect, the Game Boy Micro resembles a carefully crafted coda to the Game Boy lineage. It is a compression of the GBA’s identity rather than an expansion of it. The Micro does not attempt to be the most versatile or most practical Game Boy; instead, it distills the platform into its smallest, most focused form.

Its appeal lies partly in its refusal to behave like a sensible final model. Precise, minimal, and almost indifferent to commercial logic, the Micro compresses the Game Boy era into something immediate, tactile, intimate, and technologically self-contained. The lineage did not end with a grand summation, but with its smallest and perhaps most concentrated object.

Looking back now, the Nintendo 3DS is much more than a trick display. Inside the shell it packs an ARM11-based main system and a full DSi-class compatibility layer; it keeps DS and DSi titles alive in hardware while layering on a 3D screen, new I/O, and a very 2010s idea of “always-on connectivity.” It drifts through a whole constellation of revisions from 3DS to 2DS to the “New” models, gradually pushing 3D from headline feature to optional extra. And over more than a decade it becomes a case study in handheld hacking, moving from flashcarts and fragile entry points to boot-level exploits and custom firmware that effectively give the machine an afterlife.

This is a console built around “depth” in multiple senses: the depth on its autostereoscopic screen, and the layers of hardware, compatibility, and community work that sit behind it.

The 3D Trick: 800×240 Pixels and a Sheet of Invisible Glass

The 3D effect on the 3DS lives entirely on the top screen. On the surface it looks like any other wide 400×240 panel, but under the hood the LCD actually has 800×240 physical pixels. When 3D is enabled, the console renders two 400×240 images side by side in alternating columns; a parallax barrier in front of the LCD directs one set of columns to the left eye and the other to the right, creating depth without glasses.

That barrier is the quiet star of the show: an additional LCD layer that can form 400 dark vertical stripes. At the correct distance, those stripes obscure even-numbered columns from one eye and odd-numbered columns from the other. The result is a stereoscopic image that feels surprisingly solid if you stay inside the “sweet spot,” and instantly falls apart if you move too far off axis.

Two decisions lock the character of 3DS 3D. First, only the top screen is 3D-capable; the bottom remains a 2D resistive touchscreen. Developers are pushed into a division of labour: the upper screen becomes a “stage in depth,” while the lower screen remains a flat control surface. Second, Nintendo adds a physical 3D depth slider rather than making the feature binary. You can fade 3D in and out, from fully off to maximum strength, without ever diving into a menu. Many players eventually park the slider at zero for comfort, but the simple fact that it can be nudged back up at any moment makes 3D feel like a trick the console always has in reserve, rather than a gimmick it forces on you.

By the time the New Nintendo 3DS arrives, Nintendo patches one of the original design’s weak points: the fragile sweet spot. The “super-stable 3D” system combines the front camera and motion sensors to track your head and subtly adjust the parallax barrier in real time, making the effect hold together even if you shift a bit on the sofa. It is a retroactive fix, but it finally makes the slogan “3D without glasses” feel close to effortless in day-to-day use.

Architecture and Compatibility: Hiding a DSi Inside

Where the DS felt like a development board with features hanging off every edge, the 3DS is more stratified. At the visible top sits the 3D screen, cameras, gyro, accelerometer, wireless hardware, and NFC on later models. Just behind that, the main SoC is built around a dual-core ARM11 MPCore CPU at roughly 268 MHz paired with a Digital Media Professionals PICA200 GPU. The original models provide 128 MB of fast FCRAM and 6 MB of VRAM, while the “New” hardware doubles FCRAM to 256 MB and bumps internal caches and clocks, giving games more room to breathe.

Crucially, there is also a single-core ARM9 in the mix, a direct descendant of the DS architecture. It handles security and low-level system tasks in 3DS mode, but it also underpins the console’s backwards compatibility. When you launch a DS or DSi title, the system effectively drops into a hardware compatibility mode built around that ARM9 and the inherited subsystems from the DSi era. The extra ARM11 power and 3D hardware step aside; no DS game suddenly gains depth, and anything that depended on the old GBA slot is simply unsupported.

The result is a “console inside a console.” For native 3DS software, developers see a modern-ish handheld with a quirky GPU and two screens. For older DS and DSi games, the hardware can behave almost exactly like a late-period DS, down to offering a native-resolution mode if you hold Start or Select as you boot the title. The only major ancestor that does not survive the transition is the Game Boy Advance: there is no cartridge slot, and only a handful of GBA titles distributed in the early Ambassador Program are ever allowed to run in a special, locked-down mode.

From a systems perspective, the 3DS family is layered: at the bottom sits a DSi-class world anchored by ARM9; in the middle, an ARM11/FCRAM platform powers the main OS and 3DS games; at the top, the 3D panel, cameras, wireless, and sensors define how the machine actually gets used.

The Family Tree: 3DS, XL, 2DS, and the “New” Era

If the DS family tree was already busy, the 3DS line takes that complexity and adds more branches.

The launch Nintendo 3DS (CTR-001) is the compact original. It ships with the autostereoscopic top screen, a 2D touchscreen below, gyro and accelerometer for motion-controlled titles, a front camera and two rear cameras for 3D photos and AR games, and the now-legendary StreetPass feature that quietly logs encounters with other consoles in your bag or pocket.

The Nintendo 3DS XL/LL follows as the “big-screen” revision. Larger displays make the 3D effect more comfortable and more convincing, and the bigger shell helps battery life and ergonomics. The machine feels less like a tiny gadget and more like a living-room handheld, the sort of device you leave on a coffee table with a charger stand.

Then comes the curveball: the Nintendo 2DS. Instead of a clamshell, it has an unbroken “slab” design; instead of touting 3D, it quietly removes the feature entirely. What remains is full compatibility with 3DS and DS software, but now targeted at younger players and parents who never cared about 3D in the first place. The 2DS is an admission in hardware form: for a large part of the audience, the defining value of the 3DS ecosystem was the library, not the depth slider.

:::grid{cols=2}

:::

In 2014–2015, the “New Nintendo 3DS” and “New Nintendo 3DS XL” arrive. Internally they add faster ARM11 cores, double the FCRAM, and expand the private SRAM pools; externally they gain a small C-Stick nub, extra ZL/ZR shoulder buttons, and built-in NFC for amiibo. Certain titles, like Xenoblade Chronicles 3D and the SNES Virtual Console releases, are explicitly restricted to these models, quietly drawing a line between “can run it” and “cannot” within the same family. The “super-stable 3D” system also starts here, finally making prolonged 3D play practical.

The last word is the New Nintendo 2DS XL. It keeps the upgraded “New” internals and extra controls but once again ditches 3D, wrapping everything in a lighter clamshell shell with a friendlier price tag. By then, “3DS” as a brand means “the place where those games live” more than it means “the thing that does 3D,” and the hardware lineup quietly reflects that.

Cartridges at the Door: DS-Mode Flashcarts and Gateway 3DS

Like its predecessor, the 3DS lives alongside hacking from very early in its lifespan. At first, the activity mostly clings to the old world: traditional R4-style DS flashcarts run happily in DS compatibility mode, but they cannot see the 3DS side of the machine or its 3D hardware. They are, in practice, running on the DS that is hiding inside the system, not the 3DS proper.

The first widely known attempt to break into 3DS mode itself is Gateway 3DS. Introduced in 2013, it uses vulnerabilities in specific firmware versions to gain code execution in 3DS mode, presenting a menu that lets users mount multiple 3DS game images from a special red cartridge. For a while it is the only way to run 3DS homebrew or out-of-region titles directly on hardware.

Other products like Sky3DS take a different approach, emulating legit cartridges at the hardware level and cycling through stored titles with a button press. To the system, they look like a series of individual retail carts being swapped in and out. This sidesteps many software checks but still lives entirely within Nintendo’s expected control surface. The experience tends to be clunky, version-sensitive, and tightly constrained by how close the device can mimic an actual game card.

This era serves mostly as a proof of concept: the 3DS is not an unbreakable black box, but the early tools are fragile, expensive, and permanently tied to specific firmware ranges and hardware dongles.

Taking the Keys: Software Exploits, boot9strap, and Luma3DS

The real turning point comes when researchers and hobbyists shift their focus from proprietary flashcarts to pure software exploits. Over several years, vulnerabilities are found in retail games, the web browser, the Home Menu, and even the system’s audio player. QR-code parsing in Cubic Ninja becomes the basis for ninjhax; malformed web pages power browserhax; theme handling bugs enable menuhax; specially crafted music files trigger soundhax. Each of these gives users a way to load homebrew without any special cartridge, as long as their firmware version and installed apps line up.

Once reliable user-space execution is possible, attention moves down the stack. Exploits like arm9loaderhax (A9LH) seize control of the ARM9 during the early boot process, allowing custom code to run every time the console turns on instead of only after launching a specific app or game. With the later disclosure of secrets from the 3DS boot ROM, boot9strap (B9S) supersedes A9LH, grabbing an even earlier foothold and making installation safer and more robust.

On top of this foundation sit custom firmwares, with Luma3DS becoming the de facto standard. Luma runs as a flexible patch layer over Nintendo’s OS: it removes region locks, adds better error output, supports in-game plugins, and smooths over incompatibilities, while still letting the official user interface and eShop titles function as expected. By the late 2010s and into the 2020s, a typical “hacked 3DS” is no longer defined by which flashcart it owns but by which version of Luma and boot9strap it runs.

At that point, the community’s relationship to the hardware flips. Instead of finding cracks in an opaque box, users have a reproducible way to take ownership of the machine at boot time. The official operating system becomes one layer among several, and the 3DS turns into a general-purpose handheld platform for fan translations, homebrew, emulation, and archival projects.

After the 3D Hype: A Handheld’s Second Life

In pure numbers, the 3DS family never reshaped the world the way the original Game Boy did, but selling over 75 million units across all models is hardly a failure. It weathered the first years of the smartphone game boom, playing host to everything from mainline Pokémon entries and Monster Hunter to stereoscopic remakes of Nintendo 64 classics.

Meanwhile, Nintendo slowly winds down the official side. New hardware production ceases; the eShop closes in 2023; repair services start to sunset in one region after another. By the mid-2020s, the 3DS has clearly moved from “current platform” to “legacy hardware,” its commercial life overshadowed by the Switch and whatever comes after.

Yet in another sense, it is more alive than ever. The combination of mature custom firmware, a vast back catalogue, and a hardware design resilient enough to survive a decade in bags and drawers means that the 3DS keeps finding new roles: as an emulation handheld, a portable archive of its own era, a nostalgia machine for StreetPass and Puzzle Swap, or a testbed for late-night experiments in C and ARM assembly.

The 3D headline feature has quietly faded into the background; many people now buy or revive these systems with the slider firmly at zero. But the deeper “3D” — the layered architecture, the backwards-compatibility shell, the overlapping official and unofficial ecosystems — is what gives the 3DS its staying power. Long after Nintendo has stopped manufacturing shells and boards, and long after the last official server is turned off, the slightly battered clamshells with their custom boot logos are still booting, still being patched, still being passed around.

That is the real afterlife of this handheld: not as a museum piece, but as a machine that outlived its own marketing, and then quietly became something more interesting than the slogan on its box.

This piece walks through its architecture, compatibility, model evolution, and the stranger corners of its expansion ecosystem, treating the DS as a literal fork in the road of handheld design.

Architecture: Dual ARM and Dual Screens in Concert

After experimenting with ARM on the GBA, Nintendo doubled down on that partnership for the DS. Inside the SoC live two CPUs: an ARM9 and an ARM7. The ARM9 handles most of the DS-native game logic and 3D graphics, while the ARM7 focuses on audio, I/O, and, in GBA compatibility mode, essentially becomes the “brain” of the previous generation. It is a very Nintendo kind of compromise: not radical in raw specs, but heavily optimized for the exact use cases they cared about.

On paper, 4 MB of main RAM looks almost comically small today. But the entire rendering pipeline is designed around one core question: how do you keep two screens fed at once? Both panels are 256×192; the top behaves as a conventional “main display,” while the bottom adds resistive touch on top. The graphics engine is a hybrid 2D/3D design that can split layers between the two displays. Many games stage the “world” on the top screen and treat the bottom as a control panel; others stitch them together visually, using both screens as one tall canvas, with HUD and scene transitions bridging the seam.

For emulators, the trouble is not so much performance as entanglement. The timing relationship between the two CPUs is baked into many games. Touch input, microphone, Wi-Fi, and, in some cases, Slot-2 accessories all get wired directly into gameplay. Mainstream DS emulators can run most titles smoothly today, but reproducing the exact “feel” of the original hardware is more like simulating a whole interaction system than merely drawing frames and reading button presses. The DS is less a “faster GBA” and more a compact, oddly specialized dual-CPU interactive device.

Compatibility: The Last Window onto the GBA Era

During the “Project Nitro” phase, Nintendo set a clear constraint: the new handheld was not allowed to abandon the entire GBA software library. That decision manifested physically as Slot-2, the full-size GBA cartridge slot that occupies the entire bottom edge of the system.

Slot-2 gives the DS two layers of compatibility. The obvious one is direct GBA playback. On boot, the user can choose to run in DS mode or GBA mode. Once in GBA mode, the ARM9 more or less steps aside and the ARM7 takes over in a configuration that closely resembles a GBA. Only one screen is used, and the button layout tracks the older handheld closely. For a long time, the DS simply served as “a GBA SP with a backlight that finally did not feel like a compromise.”

The second layer is more subtle. A subset of DS titles explicitly target Slot-2 accessories: rumble modules, RAM expansion packs, and stranger, game-specific hardware add-ons. Electrically, they reuse the GBA form factor and interface, but in software they belong entirely to the DS era. This also means that when Nintendo removed Slot-2 on the DSi, these Slot-2-aware DS titles lost part of their functionality overnight. The backward bridge into the GBA world and the forward bridge into expansion hardware were cut in one stroke.

In the broader history of Nintendo handhelds, the DS is the last machine that truly “bridges” generations in hardware. It hauls the entire GBA ecosystem along, while quietly planting seeds for the 3DS architecture. The visible, literal bridge between eras is that slot that so many of us used to fill with all kinds of weird plastic.

Model Evolution: From “Brick” to Living-Room XL

The DS family went through several iterations, each answering slightly different questions about who the machine was really for.

The original NDS (NTR-001), often nicknamed the “phat” or “brick” DS, looks and feels almost like a prototype that escaped into retail. The shell is thick, the corners are sharp, and the hinge and LED layout have a somewhat utilitarian vibe. The screens are dim by later standards, and the plastic feels more functional than refined. Yet all the essential ideas are there: dual ARM CPUs, dual screens, touch input, microphone, Wi-Fi, and the Slot-1 / Slot-2 combo. It is what you get when you ship the concept before fully polishing the product.

Nintendo DS Lite (USG-001) is where that concept finally becomes a consumer device. The shell gets thinner and lighter; the hinge and edges are rounded; the backlight grows much brighter. The iconic white DS Lite, with its glossy shell and vivid screens, became the mental image of “a DS” for many people. Slot-2 receives a flush cover, so the bottom edge is clean when no cartridge is inserted, and a GBA cartridge only protrudes slightly. DS Lite preserves full backward compatibility while striking a very careful balance between “toy” and “consumer electronics.”

Nintendo DSi (TWL-001) marks a clear pivot toward the networked era. The CPU is clocked higher, the RAM jumps to 16 MB, and Slot-2 disappears, replaced by SD storage and the DSiWare download service. Two low-resolution cameras on the shell turn “playing with the camera” into a game mechanic in its own right. Simple filters, doodling tools, and gimmick apps made the DSi feel as much like a family toy camera as a traditional handheld. The trade-off is obvious: GBA compatibility evaporates, and a range of Slot-2-dependent DS titles and accessories lose their full functionality.

On top of the DSi sits the DSi XL (known as LL in Japan). It scales the formula up with 4.2-inch screens, wider viewing angles, and larger fonts. The body becomes heavier and more substantial, but the buttons and stylus feel better suited for long sessions. Pre-installed software leans toward brain-training games and party-friendly apps, positioning the machine more on the living-room table than in a student’s backpack. By this point, the DS is no longer at the leading edge technologically, but its role is clearer: not a “tiny tech gadget,” but a shared digital notebook that can be passed around the room.

The Official Expansion Zoo

Even if you ignore all peripherals, the DS is already an experimental design. The moment you start looking at the official add-ons, it begins to resemble a catalog of hardware ideas that might have been.

The Rumble Pak is the first Slot-2 accessory many players ever saw. It is a GBA-sized cartridge that lives in the bottom slot and provides simple rumble feedback in supported games. When the player takes damage, hits a wall, or triggers an explosion, the handheld buzzes briefly. Because the original DS and the DS Lite have different thicknesses, Nintendo even produced slightly different shells to keep the module from sticking out too awkwardly. The rumble is coarse by modern standards, but for a handheld of its era it was unusually “console-like.”

The Memory Expansion Pak feels like the most “engineering-driven” of the bunch. To make the Opera browser barely usable on DS hardware, Nintendo shipped this Slot-2 cartridge that contains essentially one thing: 8 MB of RAM. Combined with the 4 MB built in, the browser and a handful of later applications can stretch out over a full 12 MB of memory. For ordinary users, it is a “browser cartridge.” For homebrew developers and emulator authors, it is an extra pile of fast RAM that can be detected and exploited. Many homebrew apps probe for it on launch and quietly enable higher-resolution assets or bigger caches when it is present.

Only released in Japan, the Digital TV Tuner for DS is a time capsule from the mobile television era. It receives 1seg digital broadcasts, playing the TV feed on the top screen while the bottom screen becomes a channel guide and remote interface. A small folding antenna extends directly from the cartridge shell, snapped up when needed and tucked away when you are done. Before smartphones fully absorbed mobile video, this fantasy of “watching TV on your handheld” was surprisingly attractive. Today, the accessory reads like a fossil from a very specific moment in consumer electronics.

Among the odder peripherals, the Paddle Controller stands out as the most faithful throwback to the arcade age. It is a Slot-2 cartridge with a physical rotary knob protruding from it, designed for titles like Arkanoid DS and Space Invaders Extreme. The player uses the dial instead of a D-pad or stick, getting a direct, continuous mapping from finger motion to on-screen position. Recreating that particular feel on a tiny handheld, at a time when most controllers were going fully digital, feels almost perversely anachronistic — and that is exactly what makes it interesting.

Around these pillars cluster camera modules, guitar grips, heart-rate sensors, and various other niche devices. Most of them were short-lived. Taken together, though, they sketch out a hardware ecosystem with a particular attitude: if someone can imagine using it in a game, there is probably a piece of plastic we can build to plug into Slot-2 and make it real.

Flashcarts, Extra RAM, and the Unofficial Parallel Universe

Outside the official catalog, the DS has a parallel life built around flashcarts and homebrew.

From early R4 and M3 units to later EZ-Flash products, Slot-1 was, for a long time, the entry point for running homebrew, emulators, and various utility tools. For technically inclined users, the DS was not just a handheld for retail cartridges. It was a portable dual-ARM development platform with two screens and decent audio. With Wi-Fi and filesystem access, people built FTP clients, chat programs, music players, and even bare-bones text readers that turned the DS into a tiny, multi-purpose computer.

Some of the more advanced flashcarts also extended into Slot-2, embedding extra RAM or rumble modules there and providing software switches to toggle between “GBA link,” “rumble feedback,” and “RAM expansion” modes. In doing so, they mirrored and extended the official Memory Expansion Pak and Rumble Pak, rather than merely circumventing them. The homebrew community went as far as writing libraries and documentation specifically for these unofficial expansions, letting homebrew software stretch the DS far beyond its base 4 MB.

Viewed from today, this makes the DS look less like a closed appliance and more like an underestimated pocket computing platform. The dual ARM CPUs, programmable graphics, and rich I/O make it an ideal playground for experiments that would normally have lived only on PCs — just compressed into a clamshell that fits into a jacket pocket.

A Dual-Screen Testbed and Its Aftershocks

Seen from this distance, what kind of machine was the Nintendo DS?

Commercially, it is one of the few handhelds that can stand next to the Game Boy line in raw sales. In hardware and interaction history, it feels more like a side branch that grew away from the main trunk. If you look only at conventional metrics — CPU power, polygon count, media capabilities — the DS simply cannot trade blows with Sony’s PSP. But the DS does not try to. Instead, it builds an entire stack of interaction layers: two screens, touch, microphone, local wireless, and a flexible expansion slot, then asks developers and players to explore what that stack can actually do.

That side branch continued into the 3DS and, in a more awkward way, the Wii U’s asymmetric dual-screen experiments. It later folded back into the Nintendo Switch, where true dual screens vanished but the underlying philosophy — handheld plus dock plus detachable controllers — still carries an echo of the DS era’s obsession with new ways to interact.

When you flip open a DS today, the jagged pixels and aging plastic are hard to ignore. Yet once you remember all the Slot-2 carts, flashcarts, RAM packs, tuners, and dials that once lived around it, it becomes difficult to see it only as “the thing that came after the GBA.” It is better understood as the physical remains of a long, collective experiment in what a game console could be. The slightly yellowed clamshell in your hand is not just a product of that experiment; it is, in many ways, its most successful and least repeatable outcome.

《聲之形》這部電影，是我從未喜歡過，但卻又觀看了數遍的電影。還記得最初觀看《聲之形》是在初中，儘管當時我的思想還並不深刻，不能完全理解電影所要傳達的內容，但這部電影的劇情給當時的我造成了很大的衝擊。而最近一次觀看則是在我的高中時期，當時我已度過了自己的大低谷時期，從霸凌中逐漸走出，對於這部電影的感觸也多了一些。而時至今日，讓我再回顧這一篇電影，我會認為這是一個試圖傳遞「情感」的失敗故事。

劇情

校園霸凌——表層現象

渴望溝通——深層實質

傳遞情感

「失敗」

閒來無事，又觀看了一遍希區考克執導的心理驚悚片《迷魂記》（Vertigo）。

坦白來講，這部電影是我最愛的希區考克作品之一。一方面是這部電影的表現手段十分高明，很典型地反映出了德國表現主義對於希區考克的影響，構圖、色彩十分考究，利用了誇張、非自然的佈景與表演風格進行精神層面的表達，利用一個角色的主觀視角來理解一個奇異的世界。另一方面，這一部電影的所表達的內容也十分深刻，希區考克借助自己特有的的詭異而迷人，情感慾望交織的影像風格，塑造了「被凝視的客體」與「凝視的主體」形象，引發了對於「自我概念」「自我認知」、「男女性別」「權力控制」等議題的探討。

《迷魂記》這一部作品，在我看來，可以被解讀成希區考克的一次自省。他在前半段塑造了一位典型的希區考克式的金髮女郎瑪德蓮·艾爾絲特。而在後半段他又創造了另一位棕髮的角色茱蒂，茱蒂看似通過自己的選擇主導了後半部分的劇情，但最終還是在史考提的控制下，讓步妥協，接受了他人的塑造，又成為了另一個標誌性的希區考克式金髮女郎。茱蒂夾雜了希區考克本人略帶扭曲和邊緣化的心理特徵，在被塑造之中一步步走向絕境，最終墜樓而亡。

由於是首次租房，我並沒有什麼經驗，因此斟酌許久，選擇給自己租一間公寓。在綜合考慮租金、地段以及裝修等因素後，我最終選擇了江北區的一家公寓。從公寓出發，可以很快地來到重慶大學的沙坪壩校區，上課通勤比較方便。同時前往MAO、VOX和寅派動力等Livehouse都不算太遠，閒時可以去看演出。此外其距離幾個商圈也只需要半小時左右，其他的購物娛樂需求也能滿足。

於是，我週二我去看了房子，週五簽下了合同，週六叫了貨拉拉把大部分東西都搬到了我的公寓。又添置了許多物件，包括一些床上用品，和一把人體工學椅。在虎溪這邊忙完最後的一些工作後，下週六我便正式搬到那邊，正式開始自己的獨居生活。

如何優雅地證明 $e >\sqrt{7} $

我的解法

看到這個題目，我的第一反應是借助 $ e $ 的級數展開式 $ e = \sum^{\infty}_{n=0} \frac{1}{n!} $ 來解答，當然，本人並非數學系學生，所以下面的內容也並不嚴謹，僅僅是一個思路。

將其展開可以得到 $ e = 1+1+\frac{1}{2!}+\frac{1}{3!}+\ldots+\frac{1}{n!} $,

取前五項，計算可得 $e > 1+1+\frac{1}{2!}+\frac{1}{3!}+\frac{1}{4!} \approx2.7083$，

又易知$ \sqrt{7} < 2.6458 $，

因為$ 2.7083>2.6458 $，則可得 $e >\sqrt{7} $。

但列出這個思路之後，我不得不承認這並不「優雅」。

優雅的解法

我查閱網路，看到了一位網友運用Maclaurin級數的證明：

我們根據Maclaurin級數易知 $ e^x = \sum_{n=0}^{\infty} \frac{x^n}{n!} = 1 + x + \frac{x^2}{2!} + \frac{x^3}{3!} + \ldots $，

在此我們取 $x=2$，可得$ e^2 = \sum_{n=0}^{\infty} \frac{2^n}{n!} = 1 + 2 + \frac{2^2}{2!} + \frac{2^3}{3!} + \ldots $

又注意到截斷前四項後可得到 $ \sum_{n=0}^{4} \frac{2^n}{n!} = 7 $，

顯然，$e^2= \sum_{n=0}^{\infty} \frac{2^n}{n!} >\sum_{n=0}^{4} \frac{2^n}{n!} = 7$，

證畢。

緊跟時事的解法

因為 $abcdefg \ldots z$，

故 $e>z$，

根據姜萍等式，$ \text{主}=6 $，

而 $z=\text{主}-\text{點}=6-0.1=5.9$，

且 $5.9^2>7$，

故 $5.9>\sqrt{7}$，

因此 $e >\sqrt{7} $，

證畢。

6月14日的時候在Apple Store的軟體上預約了Apple Vision Pro的體驗。 這款產品6月28日正式發售,但是考慮周五時學校內可能有其他安排(事後證明是我多慮了),我選擇了在6月29日,即周六去體驗。

Apple Store店面玻璃上貼著AVP的標示，店內為其專門設置了一排桌子，並使用圍欄隔離開。我在排隊進入後向店員提交預約二維碼,隨後進行前期的準備。

第一是用iPhone掃描頭部的信息，需要掃描店員提供的QR code，然後用手機的Face ID掃描頭部特徵，在掃描完成後手機上會出現一個QR code，需要將其提交給店員。其實我認為這一步似乎有些多此一舉，我以為在店員出示的QR code中完全可以帶有一些追蹤信息，用戶不需要在掃描過後又提交一遍QR code。

第二是讓店員拿走眼鏡去測一下度數，用來選擇內部處方鏡片。坦白來講，這一步還是超出我的預期的，我本以為這一項是讓用戶自行選擇，但實際上卻是使用儀器去測定，這樣可以確保使用AVP的觀感和用戶正常視覺效果保持一致。

後面的工作人員會幫忙組裝適合我的設備，而與此同時店員會交代一些基本的手勢。

在設備組裝好後，店員向我教授了如何配戴設備——用右手捏住AVP的中間，將其舉起後拉住後面的伸縮帶，將其套在頭上後調整尺寸並拉緊，隨後再調整頂部伸縮帶，使之處於視野正前方。在這一步時，如果後面的調整不到位，使用中便會感受到漏光；如果頭頂的帶子調整不到位，顯示區域便不會處在視野的中心 。對於這兩點，AVP都會給予提示，確保最佳的使用體驗。（至於我是如何發現這一點的，是因為我同時犯了這兩項錯誤）

配戴後點按按鈕開機，系統根據我的眼睛距離自動調整兩個鏡片的位置。隨後需要進行處方鏡片配對。這一步需要調出兩個二維碼，並將其舉到Vision Pro的前方，讓設備自帶的攝像頭進行掃描，再配對後系統會自動對顯示畫面進行處理，使得我看到的內容清晰。

在使用時店員會使用一台iPad同步查看AVP的顯示內容，同步地給予使用方面的指導，不過我並沒能看清他的螢幕上具體還有些什麼內容。他向我詢問想要著重體驗哪方面的功能，我表示比較關注其娛樂方面。於是他指導我觀看了宣传片裡的全景恐龍視頻，線上觀看全景視頻/普通電影。此外我還瀏覽了相冊，裡面的照片有iPhone拍攝的普通/全景照片，還有iPhone 15 Pro/AVP拍攝的立體照片/視頻。

剩下的時間我還嘗試打開Safari瀏覽網頁，並嘗試調度多個窗口。這些軟件除了交互時使用手勢，其他內容都和使用MacOS差不多。

Vision Pro內置的App store裡包含了一些中國常用的軟體。據店員介紹說目前雖然可以下載很多軟體，但大部分都還沒有適配，都是直接從MacOS或IPad OS過來，未來各家廠商會對這台設備進行專門的適配。

他沒有讓我體驗Apple TV，故不知有沒有將其功能閹割。

在體驗完成後，向店員提供郵箱，他隨後便將我的信息以郵件形式發給了我，日後如果再去體驗的話，便無需再掃描頭部、測定鏡片了。

我的感受

整體體驗下來，還是非常新奇的。

戴設備時要把劉海掀起來，為觀看體驗需要將帶子拉近，但拉緊後會覺得壓迫顴骨，並不舒適。

開機後看到現實世界中的畫面，畫面很清晰，真的會讓人忘記這是在看屏幕。不過這是建立在Apple Store店舖內光線充足的情況下，在弱光環境下起顯示情況必然會變差一些。

屏幕的視角稍微有些窄，可以注意到屏幕邊緣的黑框，對沈浸感有些影響。設備的聽感還是不錯的，能感受到其「空間感」，但礙於揚聲器尺寸，臨場感還是有所欠缺。

手勢操作準確度大概85/100，也可能是我不太熟悉的緣故，隨著體驗流程進行，在後面我的操作準確度是有所提高的，且大部分操作是符合常識的。

Vision Pro的重量不算輕，在我配戴了半小時後，頸椎會感受到有些疲憊，不過眼睛沒有酸脹感。

體驗的時候是全程坐著的，這樣對於重量的感知還是會弱化一些，並且我也不清楚移動狀態下體驗會怎麼樣。

電池包是放在桌子上的，用手掂量了一下，比我的iPhone 13 Pro Max重一些。

儘管Google Workspace等服務商都提供了完善的網域郵箱雖使用方便，但所有的服務都需要付費，為了節省開支，我放棄了使用這類服務。而在眾多服務商封禁25連線埠的今天，自建郵箱服務器也不是一件易事。

通過檢索，我發現了Resend，他們會為用戶提供免費的發信服務（免費用戶100封/天，或3000封/月），恰好符合我的需求。將Resend與Cloudflare提供的電子郵件路由功能相結合，我們便可以得到一個免費的網域郵箱。

開始前的準備

1. 一個Resend帳號

2. 一個使用了Cloudflare的DNS服務的閒置網域

3. 一個充當載體的郵箱，在此使用Gmail

配置Cloudflare電子郵件路由實現收信服務

首先，打開Cloudflare的控制台，打開目標網域，隨後點擊Email選擇卡，隨後點擊Email Routing，在Custom address的輸入框中填入想要的郵箱名稱，並在下方Destination輸入框填入載體郵箱。

隨後，根據提示配置DNS解析。

根據教程操作至此，我們的網域郵箱便具備了收信功能，如不需要發信服務，到處便已經結束。

配置Resend實現發信服務

打開Resend官網，在首頁或API Keys頁面申請一個免費的API Key，並將其複製下來備用。

在Settings頁面選擇SMTP選擇卡，獲得SMTP主機、連結埠等信息。

登錄載體郵箱，在設置介面找到「添加其他電子郵件地址」，單擊。

根據提示填入各項信息，「電子郵件地址」處填寫網域郵箱地址。

根據Resend官方文檔填寫SMTP伺服器（smtp.resend.com），選擇連結埠（465），填寫用戶名（Resend）。結果如下圖所示。

單擊「添加帳號」，將網域郵箱添加至載體郵箱帳號。隨後打開載體郵箱的收件箱，找到認證電子郵件並進行驗證。

經過上述配置，我們便能夠在載體郵箱的發信介面選擇我們的網域郵箱，並可使用其發信。

參考文獻

域名搁着别浪费，Cloudflare + Gmail + Resend 十分钟轻松拥有免费的企业邮箱

Resend Documentation

A Cross-Cultural Colour-Naming Study. Part I - Using an Unconstrained Method

A Cross-Cultural Colour-Naming Study. Part II - Using an Constrained Method

A Cross-Cultural Colour-Naming_Study. Part III - A Colour-Naming Model

Colour Naming - Linking Vision and Speech

Color Names Across Languages: Salient Colors and Term Translation in Multilingual Color Naming Models

第一次合唱比賽是在初中，曲目要求是一首「紅歌」與一首流行歌曲。那次我們獲得了年級第一名的成績。

當時我是我們班的班長，而副班長、文藝委員、組織委員都是我的好友。在得知合唱比賽的消息後，我們班鬥志昂揚，許多「猛士」都積極獻策。從他們那裡，我們選擇購買了婚禮慶典上常用的「禮炮」——一種旋轉底部氣管，釋放壓縮氣體，便可以發射出大量彩帶的紙筒——作為氣氛道具，並定下來了流行歌曲爲Queen 樂隊的We Will Rock You。

文藝委員和組織委員則主要負責班級內同學的聲樂合唱。而幾名具備器樂基礎的同學則被我每天課餘時間拉走特別訓練——兩位同學進行吉他伴奏，一名同學負責鋼琴，而我則使用貝司墊低聲。至於打擊樂組，我則邀請了一位隔壁班的外援「Z」——Z從小便學習爵士鼓，技術在本市內都算是頂尖，而由於和他們班同學不合，沒有幫助他們班，而卻選擇來幫助我們班演出。

正式比賽的場地在初中部報告廳，恰好就在我們班教室的對面，我們近水樓台先得月，一旦得閒便前去排練，相比於其他班級，我們打磨出了更適合的隊形站位，甚至還設計了在兩首歌交替之間的隊形變換。

我所使用的貝司與音響，是我和副班長從管樂團裡臨時偷來的。當時我們的校管樂團恰好正在排Deep Purple的Smoke on the Water，我正負責其中的貝司。在正式表演前一天晚上，我們將音響和貝司悄悄從學校一角的藝術樓搬到了我們的教學樓。

經過幾次排練後，我們在正式比賽那天穩定發揮，第一首「紅歌」時我擔任班級的合唱指揮。整首歌下來，除了在鞠躬時我險些摔倒，我們整首歌沒有任何瑕疵。隨著兩首歌交替，我們班的隊形站位迅速改變——當時在場的其他班級都未曾有過類似的舞台設計——我和同時器樂組的其他同學也及時操起樂器，準備等待開始。

第二首歌的過程中，我並沒有留下太多印象，只是確定我們的演出相對穩定，儘管Z事先並未和我們排練，但他憑藉自己過硬的器樂素養，還是為我們的演出錦上添花。

第二次合唱比賽發生在二〇二〇年的八月份。當時第一波武漢肺炎疫情逐漸穩定，我們六月初才返回學校。在學校進行了短暫的線下學習，並進行了期末考試後，學部便組織了這一比賽來振奮我們的精神。在這一次，我被教務處的老師臨時徵用，負責後台的調音、燈光。班級內的各項事務都交給了文藝委員負責。但在比賽前一天晚上，我們突然收到通知，每個參賽班級都需要準備背景PPT或視頻。時間緊急，文藝委員又叫來了在報告廳後台忙碌的我去幫忙。由於時間緊迫，我也只能劍走偏鋒，讓家裡安排人儘快送來拍攝設備，我在班級內快速錄製了幾條視頻素材。

但有了素材並不能解決問題。為了剪輯成片，我向老師借來了筆記本電腦，但其配置又不能滿足要求，於是我只得又遠程打開家裡的工作站，從老師的筆記本電腦遠程連結，共享素材進行剪輯。眼看天色已晚，我只能拿著筆記本儘快回到宿舍。

當我拎著電腦，回到房間，將其擺在桌子上時，宿舍執勤阿姨便立刻來到我們房間，看著我的電腦，感慨儘管很快就會放假，但現在的學生也不能這樣放肆。我的班主任恰好在我的房間，及時地欸我解圍：「小X要為我們學部剪輯一條視頻。」——這個老狐狸，為了令宿舍阿姨同意，謊報軍情，明明我的任務屬於班級，但他卻稱「爲學部剪輯」。宿舍阿姨見此狀，便不再多言，離開了我的房間。

但熄燈之後，樓長又正巧來到我們樓層巡視，並且從我們房間的門口，看到了面對著發光螢幕「夜戰」的我，她進來後，看著桌子上的筆記本電腦、無人機，以及一大瓶飲料，坐在桌前的我正看著手機打著字。樓長雖工作經驗豐富，但也未曾見過此場面。「現在的學生都這麼猖狂了嗎？這些東西沒一個是合法的！」她正準備訓斥我，我們的宿舍阿姨遲遲趕來，向樓長解釋「X同學要幫Y主任做個東西」——Y主任是我們的學部主任，脾氣火爆且及其較真，宿舍阿姨在班主任編造謊言的基礎上還加深了一部，為我分配了一位領導！

樓長聽到Y主任的大名，也不敢再多嘴，只說了一句不要影響其他同學便關上門離開。而我在這之後便通宵奮戰，最終在早晨七點時為我們班級交付了一條十分完美的視頻。

在我們的選曲中有一段RAP，當時我還特別邀請了我們班一位熱愛說唱的同學，跟著音樂對口型，而我手持鏡頭對著他的正面拍攝特寫。他左右搖擺地向前走，我便配合他的動態一起搖晃，同時謹慎的向後退，確保他始終位於畫面中央。在RAP的最後階段，他手向正排成方陣的我們班同學一指，我也旋轉鏡頭，一齊指向了我們班的同學，實現了完美的配合——這一部分是我們第一次試拍時即興發揮的鏡頭，後面補充鏡頭時曾試圖對這一段多拍幾條，但再也拍不出那種渾然天成的感覺。

在比賽當天，我們班無疑在此獲得了第一名的成績——雖然我很想開玩笑說這多虧了我在幕後的配合，但我不得不承認，這是那段時間我們班所有同學一齊努力奮鬥的結果。

在這場比賽後，我們經過了一個短暫的暑假。隨後，我們升入了高二，學校根據選科以及成績，將我們班三十人拆散到了八個不同的班級。但我們雖然不在同一班級，但關係依舊親密。學期初班幹部競選，年紀裡八個班，七個班級的班長出自我們班（剩下的那個班級，在一個月試用期後被撤職，也換上了我們班的同學擔任班長）。每次年級舉辦會議，都可以稱作我們高一班級成員的團建。

那時，我的中考成績並不是自己查詢到的，而且還比別人早了幾個小時——至今我也不明白為什麼會這樣。

記得那一天凌晨，我剛剛從上海飛到廈門。在酒店稍作歇息，早晨起床去衛生間時，一位熟識的音樂老師突然打電話給我。我茫然地接通電話，聽到他不正經地問我是否想知道自己的中考成績。我隔著電話翻了個白眼，說道：「成績公布不是還要等幾個小時嗎？你又在賣什麼關子？😤」

他聽完嘿嘿一笑，繼續問：「想不想知道嘛？」在我肯定答覆後，他便告訴了我我的中考成績。說實話，那个成績確實有些出乎意料，但我並不清楚自己在全市的排名，於是向他打聽，「其他人的成績怎麼樣？」

然而，他卻有些為難，只告訴我他只知道我的成績，不知道其他人的。

向他道謝後，我掛斷電話，把這一消息告訴了家人。父親甚至還顯得有些失望。

不久後，我的微信突然收到了大量消息，同校的老師和同學紛紛向我詢問我的中考成績。經過一個朋友的描述，我大致明白了事情的經過。

當時，我的初三班主任正帶著其他同學在日本進行由學校組織的畢業旅行。而那個通知我成績的音樂老師在和我通話後，立刻把消息告訴了班主任，班主任隨後將我的成績告知了一起旅行的同學們。於是，在中考成績公布前的幾個小時，幾乎我們學校的所有人都知道了我的中考成績，而僅僅是我的成績。

一位那時在日本旅行、後來成為市狀元的同學，在得知我的成績後，還小聲嘀咕：「xxx這次怎麼考得這麼高」，被市第二名向我打了小報告。看來他們也許認為我有可能成為本屆的市狀元。（雖然最終並沒有）

幾個小時後，成績才正式公布，但我和家人並沒有費力去查詢。畢竟分數已經知曉，高中也早已簽約，查詢顯得多餘了。相比於登上那個卡頓的查詢網站，我更傾向於享受我在廈門的短暫假期。

然而，我在廈門也並未怎樣享受。同一天的下午我便收到了教務處的電話，他們向我表示，希望我能夠參與組織排練學校的畢業式。我表示我正在廈門度假，婉拒了邀請，老師便又詢問我能否作為畢業生代表在典禮上進行演講。心中感到不好拒絕，便勉強答應了下來——在我答應後，老師卻告訴我畢業典禮正舉辦在三天後。

於是，我只得取消後續幾天的度假計畫，訂好第二天從廈門回家的高鐵票。在高鐵上打好演講稿的草稿並和老師進行了多次聯合修改，我終於定下了一篇還說的過去的講稿。

畢業典禮當天，我的演講差強人意。雖然在演講中提到了所謂的畢業離別，但我也清楚，我們初中的大部分同學還是會繼續在本校唸書，我所講的實際上也指示客套話。

更令人啼笑皆非的是，在畢業典禮的後一天，我們便需要來到學校，進行「銜接課程」的學習。好巧不巧，我的瑞士度假計畫正巧也訂在了「銜接課程」的開學日，於是我向老師謊稱我要去歐洲「遊學」，避開了萬惡的補課。

至於選擇更換網域的原因，主要還是我對於這個站點的定位做出了調整。我希望不再將這一個站點作為我個人在短期內的博客更新站點，而是一個長期的、穩定的中轉站。基於這一個網域，我未來會部署各類方便我與周邊朋友的網路服務，以及打造個人品牌。基於這樣的目的，使用「nishikori_tech」似乎顯得小家子氣。況且「錦織 唯」只作為此時此刻我個人的筆名，我並不能確保我未來是否會繼續沿用這一名稱。

對於網域的選擇，我也糾結了較長時間。由於個人的審美偏好，我傾向於將網域的立意定在「起源，本初」。

最適合這一立意，且符合本站以中文為語言的定位的，當數「元」的拼音「yuan」，但可惜作為中文拼音，絕大多數的「yuan」網域都已被註冊，我最鍾意的「yuan.io」雖處於代售狀態，但$3,500的最低價格讓我望而卻步。

隨後，我想到了使用「元」的耶魯式拼音「ywan」，但由於其過於冷門，因此作罷。

隨後又從各個語種考慮，在網域註冊站點上篩選了許久，查詢了「alfa」「origine」「fundementa」「kigen」「genesis」等等諸多方案，在簡潔程度、價格費用、立意貼近程度數個維度考慮後，我最終選定了如今這個網域名稱「ursrpung」——德語詞彙，通常被理解為「根源」「起源」「本源」。

希望未來的時間裡我可以持續使用這一網域名稱（除非我真的能買下更好的）。

前段時間熬夜完成了中國國際大學生創新大賽的項目申報書，希望一切順利。

數學建模比賽的結果暫時也還沒有得出，同樣希望順遂。

這個週末說實話沒怎麼認真複習，多數時間都在睡覺和刷手機。

之前常用的圖床IMGTP突然停止了服務，原有的圖片基本都無法顯示了。嘗試搶救了幾篇文章的配圖，但感覺耗時太長，決定日後有時間再來慢慢恢復。畢竟圖片都是自己的，完全恢復的可能性是有的，但是工程量相對來說還是比較大的。

前些日子，前任因為某些原因選擇輕生。還好她被及時救回來，並和我恢復了聯繫。無論當初內心有多不甘，我還是希望她能夠好好的。

五一假期沒有機會出去玩，因為我需要認真備考托福，我的家人不支持我外出旅遊。整個假期總結來說也並不愉快，不過同時也開始了一些自學的計畫，目前在Coursera學習Meta的前端課程（包括React.js的基礎），期待以後能夠自行維護更新這一博客站點。

4月30日去看了宮崎駿《霍爾的移動城堡》的重映，以及《SPY×FAMILY間諜家家酒》的電影《劇場版 SPY×FAMILY CODE: White》。後者我觀看的是首映特殊IMAX場，因此我有幸拿到了限定海報。

一個月裡陸續做了許多材料力學的實驗。讓我印象比較深的就是材料的壓縮、扭轉實驗——雖然一直以來對於這幾個材料的性質爛熟於心，但真當我在實驗中見識到這些材料究竟能經受多大的壓縮/扭轉之後，還是不禁驚嘆出來。

:::grid {cols=2,gap=12}

:::

去了過去兩年一直很喜歡的店鋪吃了壽喜燒，一直以來很喜歡的黃油鍋底下架了，肉的種類也變少了，估計以後不會再去了。

托福考試整體來說還是比較順利的，大概是可以和托福分手了。

托朋友買來了李志在日本巡演時售賣的專輯。不得不感嘆現在在國內想買到一張逼哥的專輯真的很難——其實說實話，在中國買專輯本身也不是一件容易事了。週末有閑，在渝中區閒逛，去了兩家音像店。一家如今十幾年來搬家數次，雖仍在營業，但店內專輯價格大概是正常價格的兩倍。另一家音像店卻不知什麼時候已經倒閉，我費盡力氣來到它的「遺址」，卻只得看到門口堆積的落滿灰塵的貨架。

拿著新的托福成績單，以及英語學分兌換申請表，來到了沙坪壩校區，想讓我們學院的教務老師簽字蓋章。但前日的留言她尚未回覆，辦公室裡她的工位又不見人影。一氣之下我住在了沙坪壩校區的附近，又跟老師發了信息。隨後正當我在社交平台上吐槽時，這位教務老師卻戲劇性的回覆了我，告訴我只要符合條件，便可以直接去找她蓋章——好似這是一件十分容易而順利的事情。雖然滿是怒氣，但我卻只能順從地回覆一句「好的」。

那天我吃了麥當勞，20元20個雞塊，與我在香港時點過的套餐相似，但中國的麥當勞的包裝設計著實更好一些。

剛剛過去的這一週，我與隊友忙於數學建模校級競賽，晝夜不分地趕製論文，終於在截止日期前提交上去。就我個人感覺，這篇文章算是完美的。希望一切順利。接下來的時間我還需要負責撰寫大創比賽的申報書，又要透支自己的身體。

星巴克最近推出了挺多的新品，個人最喜歡第三張圖片的「紅顏草莓椰椰星冰樂」。

::: grid {cols=3,gap=4}

:::

友人前幾日向我發出邀約，邀請我一同去韓國休假。儘管我的假期任務繁多，但考慮到這個學期我並未好好度過任何一個假期，我還是想在暑假找個時間放鬆一下。在當日我便與母親商量，她欣然同意，也說我這個學期都沒休息過，可以在假期裡抽出一點時間放鬆一下。

然而，當她昨天下午將這件事告訴我的父親時，他卻直接否決。我的母親雖表面上向著我，但實際上並不會忤逆我的父親，向我傳達了父親的想法。昨晚我們通電話時，他「苦口婆心」地告訴我我的暑假任務很多，需要把精力放在「正事」上面。

他向來都是這樣，一個典型的中國家長，一邊標榜著自己「民主」，一面又不容他人質疑，獨斷專行，像極了偉大的中國共產黨。當我們之間發生一個問題時，他從不會說「因為現在這個情況，讓我們想想怎麼做」，而是「因為現在這個情況，你應該這樣做」，看似留給了我選擇的空間，但其實他已然「幫我」做好了決定。若我對他反駁，他便會說出那經典的「還不是為了你好」之類的言論，似乎我反駁他，做出自己的決定如同十惡不赦的大錯誤。

我懶得和他繼續爭辯，但這也不代表我永遠會忍氣吞聲。當我真正地不再依附於他時，能夠做出自己的選擇時，我不會再顧及他任何。

今天晚上下著雨，但我並沒有帶傘，從星巴克一路淋著雨回到了宿舍。在路上我和他又通了電話，期間都是虛偽的和諧，我向他報告著近幾天學習、工作的情況，他也象徵性的關心了一下我淋雨會不會感冒。比起我的精神狀態，我能夠不生病，能夠繼續投入學習才是更重要的。

從小我便是所謂的「別人家孩子」，有關學習的事情從來沒有讓他操過心——當時他調任外地，也不怎麼關心。我輾轉了兩個南方城市才讀完了小學，直到上了中學才穩定居住在山東，與他們分居。中學是住宿制，我每週或每半個月離開學校一次，獨自搭城際公交回到只有自己的家。偶爾我會因為感情問題（中國所謂的「早戀」）被老師通知家長，我的父親也最多只是在電話裡說我兩句。

他真正開始「關心」我，是在他回到了山東，我經歷了嚴重的心理問題之後——我的心理問題是靠自己慢慢解決的，在我逐步回復正常後，我的考試成績經歷了嚴重的下滑。是的，直到我成績沒有了當初的光鮮後，我的偉大的父親才開始關心起我來。當然，他的關心並不是那麼的體貼，而是囑咐我認真學習，不要分心，把成績提上去之類的隔靴搔癢般的話，而從來沒有關心過我的精神狀態。我猜，直到今日，他可能都不清楚我曾經經歷過一段十分痛苦的時期。

在我獨自排解了自己的心理問題後，我逐漸接納了自己，認識到了自己只是一個凡人，也不再對自己有任何不切實際的幻想。但在他的眼中，我這是放棄了奮鬥和努力，我作為「他最好的作品」（他依舊把我當作他的「附屬物」，而不是獨立的「人」），不應該如此的消極（儘管事實並非如此）。

高考完之後，我對自己的成績並沒有什麼後悔，我接受現在的我——但他不允許。他讓司機帶著我連夜來到長沙，去和國防科技大學的某位熟識的政委「協調」，試圖將我送入這所在他看來十分不錯的學校。但我不肯，因為我深知倘若真的進入了這所學校，我的未來就消失了。我在國防科技大學的校門前和他吵了起來，最終讓他改變了想法，讓我按照常規的路徑升入一所普通的大學。我在報考志願時，他一直指手畫腳，保留下來的都是他所中意的大學，直到填報到最後還剩下幾個空餘時，我才有機會在裡面加上了自己的意願學校——其中幾項志願便是重慶大學。其實當時我對於重慶大學一無所知，唯一的了解便是這個地方距離我的家很遠。

我最終被重大錄取，來到了這個距離山東近兩千公里的城市。這是我對偉大的父親的第一次勝利，也是至今唯一一次。

管理決策分析就是為幫助決策者在多變的環境條件下，進行正確決策提供的一套推理方法、邏輯步驟和具體技術，以及利用這些技術和方法選擇滿意行動方案的過程。

決策分析類型

決策問題的基本模式：

• $ A_i $ 是決策者第$i$種策略或方案，屬於決策變量，是決策者可控因素；

• $ \theta_j $ 是決策者和決策對象（決策問題）所處的第$ j $種環境條件或自然狀態，屬於狀態變量，屬於狀態變量，是決策者不可控制因素；

• $ W_{ij} $是決策者在第$ j $種環境條件下選擇第$i$種策略的結果，是決策問題的價值函數值，一般叫益損值、效用值。

確定型問題分析

問題特徵

1. 存在決策者希望達到的目標（收益大火損失小）；

2. 存在著可供選擇的兩個以上的行動方案；

3. 不同行動方案再自然狀態下的益損值可以計算出來；

4. 存在一種確定的自然狀態；

求解方法

1. 當方案數量較大時，常運用運籌學中規劃等方法來分析解決（線性規劃、動態規劃目標規劃等）；

2. 如多階段確定型決策問題的求解方法——動態規劃方法；

3. 嚴格來講，確定型問題知識優化計算的問題，而不是真正的管理決策分析問題。

典型例題

某企業準備生產甲產品，其單位售價為150元/件，單位可變成本為100元/件，固定成本為5000元，年產量200件。求解：

1. 企業年獲利多少？
2. 盈虧平衡產量為多少？
3. 最低定價為多少時，企業不會虧損？
4. 若目標利潤為1萬元，目標成本為多少？
5. 原材料上漲和工人工資增加，使得單位可變成本上升至140元/件，若單位售價不變且無法轉產，企業是否該停止生產？

企業年獲利：

盈虧平衡產量：

盈虧平衡產量：

目標利潤為1萬元的銷售量：

目標總成本：

風險型問題分析

問題特徵

1. 存在決策者希望達到的目標（收益大或損失小）；

2. 存在著可供選擇的兩個以上的行動方案；

3. 不同行動方案在自然狀態下的益損值可以計算出來；

4. 存在兩種或以上的自然狀態；

5. 可以預測不同自然狀態發生的概率。

求解方法

1. 期望值；

2. 矩陣法；

3. 決策樹法。

風險型決策分析問題是一般現實情境中決策分析的主要內容，在基本方法基礎上，應注意把握信息價值及其分析以及決策者效用觀等重要問題。

不確定型問題分析

問題特徵

1. 存在決策者希望達到的目標（收益最大或損失最小）；

2. 存在兩個以上可供選擇的行動方案；

3. 不同行動方案在自然狀態下的益損值可以計算出來；

4. 存在兩種或以上的自然狀態；

5. 不同自然狀態發生的概率無法預測。

求解方法

1. 樂觀法（最大最大原則）;

2. 悲觀法（最小最大原則）;

3. 後悔值法（SaVage準則或後悔值最大最小原則）;

4. 等概率法（Laplace准則，一種特殊的風險型決策）。

典型例題

某企業準備生產一種新產品。估計該產品的銷售量有較高、一般、較低、很低四種情況，而對每種狀態出現的概率則無法預測。為生產該產品，企業有三種實施方案：新建一個車間進行生產；改造一個現有車間進行生產；部分零件在現有車間生產，部分零件外購。該新產品企業準備生產10年，10年內在不同狀態下的損益值（扣除投資費用）如表5-1所示。請分別用樂觀法、悲觀法和後悔值法來決策實施方案。

樂觀法：

每個不同方案在不同狀態下的最大受益分別為：

取各方案最大受益值中的最大值，可得：

即對應方案$ A_1 $，建立新車間。

悲觀法：

每個不同方案在不同狀態下的最小受益分別為：

取各方案最小受益值中的最大值，可得：

即對應方案$ A_3 $，部分生產、部分外購。

後悔值法：

取每個狀態下最大的收益值，將之與其餘方案下的收益值做差，隨後求的每個方案的最大後悔值進行比較。

取各方案最大後悔值中的最小值，可得：

即對應方案$ A_2 $，改造現有車間。

等概率法：

假定各狀態發生概率相同，通過計算各方案的收益期望值進行比較，取個方案收益期望值的最大值，選擇對應方案，因方法過於簡單，不再贅述。

風險型問題分析

問題特徵

1. 存在決策者希望達到的目標（收益大或損失小）；

2. 存在著可供選擇的兩個以上的行動方案；

3. 不同行動方案在自然狀態下的益損值可以計算出來；

4. 存在兩種或以上的自然狀態；

5. 可以預測不同自然狀態發生的概率。

求解方法

1. 期望值；

2. 矩陣法；

3. 決策樹法。

風險型決策問題是一般決策分析的主要內容，在基本方法的基礎上，應注意把握信息價值及其分析以及決策者效用觀等重要問題。

期望值法

期望值法就是利用概率論中隨機變量的數學期望公式算出每個行動方案的益損期望值並加以比較。若採用決策目標（準則）是期望收益最大，則選擇收益期望值最大的行動方案為最優方案；反之，若決策目標是期望費用最小，則採用費用期望值最小的方案為最優方案。

• $ X_i $是隨機離散變量$ X $的第$ i $個取值，$i=1,2,\cdots,m $;

• $ P_i $是$ X = X_i $時的概率。

決策樹法

所謂決策樹法，就是利用樹形圖模型來描述決策分析問題，並直接在決策樹圖上進行決策分析其決策目標（準則），也可以是益損期望值或經過變換的其他指標值。

典型例題：

某輕工企業要決定一產品明年的產量，以便及早做好生產前的各項準備工作。假設產量的大小主要根據該產品的銷售價格好壞而定。根據以往市場銷售價格統計資料及市場預測信息得知：未來產品銷售價格出現上漲、價格不變和價格下跌三種狀態的概率分別為0.3，0.6和0.1。若該產品按大、中、小三種不同批量（即三種不同方案）投產，則下一年度在不同價格狀態下的益損值可以估算出來，如下表所示。現要求通過決策分析來確定下一年度的產量，使該產品能獲得的收益期望為最大。

根據題目描述可繪製出決策樹如下

多級決策樹法

如果只需作一次決策，其分析求解即告完成，則這種決策分析問題就稱為單級決策。反之，有些決策問題需要經過多次決策才告完成，則這種決策問題就是多級決策問題。應用決策樹法進行多級決策分析稱為多級決策樹法。

信息的價值

信息與決策的關係十分密切。要獲得正確的決策，必須依賴足夠和可靠的信息。決策所需信息的分類：一類是完全信息，即據此可以得到完全肯定的自然狀態，有助於正確的決策；一類是抽樣信息，這是一類不完全可靠的信息。

完全信息價值

典型例題

某化工廠生產一種化工產品。對統計資料的分析表明，該產品的次品率可以分成五個多級（即五種狀態），每個等級（狀態）的概率如下：

由進一步的分析發現，產品次品率的高低與該產品所用主要原料的純度有關。今已知，化工原料純度高次品率低（如$ S_1 $為 0.02），反之則次品率高。而化工原料的純度高低，又與運輸、保存日期等因素有關。潮此，工廠主管生產的部門建議在生產該產品前，先對該化工原料增加一道「提純」工序，通過提純工序，能使全部原料處於$ S_1 $狀態，從而降低了次品率。但增加提純工序就必須增加工序費用。

經過核算可知，每批原料的提純費用為3400元。經估算，在不同純度狀態下其益損值如下表所示。如果在生產前，先將化工原料檢驗一下，通過檢驗可以掌握每批化工原料處於何種純度狀態，這樣可以對不同純度的原料採用不同策略，即提純或不提純，從而使益損期望值為最大。

根據題目描述可繪製出決策樹如下

抽樣信息價值

典型例題

某公司有50,000元多餘資金，如用於某項目開發，估計成功率為96%，成功時一年可獲利12%，但一旦失敗，有喪失全部資金的危險。如把資金存放到銀行中，則可穩得年利6%。 獲取更多情報，該公司求助於諮詢服務，諮詢費用為500元，但諮詢意見只提供參考。根據過去諮詢公司類似200例諮詢意見實施結果，具體情況見下表所示。

根據題目描述可繪製出決策樹如下

首先，我們要明確，二手玫瑰參加這一節目，並不是他們主動向《歌手》節目組請纓，乃至腆著臉申請參加節目。相反，他們的登台必定是節目組的主動邀請。當節目組發出邀請時，我猜測二手玫瑰內部也會存在顧慮——他們作為經驗豐富的樂隊，必然了解自己的音樂風格並不會被主流人群全然接受，即使是他們的樂迷，也會發表「正常人誰聽二手玫瑰啊」「二手玫瑰是中國最大的紅白喜事樂隊」的聲音。何況他們作為一個樂隊，相比於中國電視觀眾喜聞樂見的「炫技」，他們的作品更突出的特點在於情感的抒發。

但是既然節目組對他們發出了邀請，他們還是欣然赴了約。节目组邀请他们，大概率是想體現節目的包容性，以及表達對各類喜好的觀眾的關注，或者，功利性地說一句，這是為了增加這一節目的熱度。

但是在《歌手》上的演出，帶給一些人反饋卻是「這怎麼這麼難聽？」這個問題我認為實際上要歸咎於節目組。在流行歌曲中，通常採用人聲聲音主導，伴奏在背景支撐的混音方式。而在樂隊作品或者人聲樂器化的作品中，人聲和樂器更為接近。這兩種不同類型的作品有不同的混音審美取向。

然而，在《歌手》的名字我們便可以看出，其側重點在「歌手」而不是「樂隊」，這意味著節目的混音重心在於歌手的人聲。因此，製作組在混音二手玫瑰的音頻時仍然按照流行音樂的混音規則，將人聲混在器樂之上並佔據主導地位，即人聲浮於器樂之上。這可能會導致演奏中的打擊樂和吉他部分不那麼突出，換言之，將這樣一首渾然天成的，作為一個整體的樂曲給拆散了。線上觀眾們做出「難聽」的評價也無可厚非——甚至是熟悉搖滾樂的我，也想做出這樣的評價，認為這場表演並不如《樂隊的夏天》中同首歌的演繹。

為什麼我認為問題出在「混音」，這是因為在其表演中，現場觀眾的反饋實際上是不錯的，這暗示我們在現場存在自然混音，且音響聲壓足夠時，這首歌的聽感還是不錯的。

再來談一下我對二手玫瑰這個樂隊的看法。中國搖滾樂的歷史實際上是本土創作和對國外樂隊的模仿相結合的產物。一些自稱「具有國際視野、時尚風格」的樂隊，並沒有呈現出與眾不同的特點。許多人仍未意識到，民族特色才是獨具魅力的。在從事音樂創作時，至少應該有自己獨特的想法。我希望能夠聽到更多體現我們國家民族特色的音樂作品，因為我們擁有獨特的文化傳統，亟待傳承和發揚。在這些年里，我特別喜歡的幾支具有中國特色的樂隊包括二手玫瑰、九寶和杭蓋等，他們在演奏中融入了中國的本土樂器（如嗩吶、大鼓等），在增加了音樂厚度的同時也讓更好成就了中國搖滾的「本土化」。

剛剛發現《歌手 2024》的 EP2 已經放送，他們選擇了《嫂子頌》。這一首歌的原唱是李娜，歌曲內容十分悲壯，而二手玫瑰對這首歌的演繹並沒有給出異樣感（即作為一個男歌手，梁龍並沒在演唱中表現出油膩/反差感）。

他們在這場比賽上的選曲是別出心裁的，沒有選擇例如《仙兒》之類相對更火，流傳度更廣，更容易被他人接受的歌曲，而是兩首有態度，有表達的歌曲。

我認為這表達了他們參加這個節目的態度——二手玫瑰不去迎合觀眾，而是去表達自己真正想表達的深沈的東西。

關聯矩陣法

關聯矩陣是一種常規的綜合評價方法，它採用矩陣的形式，表示各備選方案針對具體指標價值評定量之間的關係。

• $ A_1 , ⋯ , A_m $是某評價對象的$ m $個替代方案；

• $ X_1 , ⋯ , X_n $是評價替代方案的$ n $個評價指標或評價項目；

• $ W_1 , ⋯ , W_n $是$ n $個評價指標的權重；

• $ V_{i1} , ⋯ , V_{mn} $第$ i $個替代方案$ A_i $關於$ X_j $指標$ ( 1 \leq i \leq m, 1 \leq j \leq n ) $的價值評定量。

下面將結合一例題來接受兩種確定權重以及價值評定量的方法。

某企業生產一款熱門產品，規劃了以下三種生產方案：

$ A_1 $：自行設計一條新的生產線；

$ A_2 $：從國外引進一條自動化程度較高的生產線；

$ A_3 $：在既有設備基礎上改裝一條生產線。

經過權威部門及專家討論，評價指標確定為五項，分別為：期望利潤；產品良品率；市場佔有率；投資費用；產品外觀。

根據專業人士的預測和估計，執行這三種方案後，對五個評價項目的結果如下所示。

逐對比較法

這一方法的基本步驟為：

1. 對各替代方案的評價指標進行逐隊比較，對於相對重要的指標給予較高得分，得到各評價項目的權重$ W_j $；

2. 根據評價主體給定的評價尺度，對各替代方案在不同評價指標下一一評價，得到相應的評價值；

3. 加權求和得到綜合評價值。

對於例題，我們首先需要通過逐對比較法求處個評價指標的權重。如下圖所示，期望利潤與產品成品率相比前者更為重要，故前者得一分後者得零分。最後根據各評價項目累積得分計算權重。

隨後，由評價主體確定評價尺度，如下圖所示，以使方案在不同指標下的事實結果能夠統一度量，以便加權求和。

隨後根據前兩張圖，對各替代方案進行綜合評定，結果如下圖所示，可知$ V_2 \gt V_1 \gt V_3 $，則$ A_2 \gt A_1 \gt A_3 $。

古林法

但對各評價項目間的重要性可以做出定量估計時，可以採用古林法，這是確定指標權重和方案值評定量的基本方法。

這一方法的基本步驟為：

1. 確定評價項目$ R_j $的重要程度；

2. 將$ R_j $標準化，以$ K_j $為單位處置，並得到權重$ W_j $；

3. 採用關聯矩陣法相同的方式計算替代方案的評估值；

4. 加權求和得到綜合評價值。

對於例題，我們需要首先確定評價指標的重要度$ R_j $。

隨後對$ R_j $進行基準化處理，設基準化處理的結果為$ K_j $，以最後一個評價指標為基準，令其$ K $值為1，自下而上計算其他項目的$ K $值。

在完成上面的步驟後，對$ K_j $進行歸一化處理，將$ K_j $列數字相加得到$ \sum K_j $，如圖所示用$ K_j $除以$ \sum K_j $，所得結果紀為$ W_j $。

算出各評價項目的權重後，按照同樣的計算方法，對各替代方案逐項進行評價。此處方案$ A_i $在指標$ X_j $下得重要度$ R_{ij} $不需要再予以估計，可以按照各替代方案的預計結果按比例計算出來。如對期望利潤$ X_1 $的$ R_{i1} $，因$ A_1 $的期望利潤為650萬元，$ A_2 $的期望利潤為730萬元，則有$ R_{11} = 650/730 = 0.890 $，$ R_{21} = 730/520 = 1.404 $，隨後按照與第二步相似的方法求得$ K_{ij} $並進行歸一化得到$ V_{ij} $。

在計算第四步時，由於投資費用越小越好，故其比例求取倒數，即$ R_{14} = 180/110 = 1.636 $，$ R_{24} = 50/180 = 0.279　 $

可知$ V_2 \gt V_1 \gt V_3 $，則$ A_2 \gt A_1 \gt A_3 $。

簡便方法：

將所有替代方案中的各項價值評定量除以最後一行的值；

隨後將各行的值除以各列的值的總和；

進行歸一化，便可以快速得出加權和$ V_i $以得出結果。

層次分析法

使用層次分析法建模大致可依四個步驟進行：

1. 建立層次結構模型，描述評價系統中各元素之間的關係；

2. 創建兩兩判斷矩陣，對同一層元素進行兩兩比較，構建矩陣；

3. 計算要素相對權重，分析同一層每個元素對上一層準則的重要程度，透過判斷矩陣計算各元素相對權重，並進行一致性檢驗；

4. 計算各元素對系統總目標的合成（總體）權重，並對各候選方案進行排序。

在此使用例題進行說明。

某大学需要科学评审一系列科研课题，选择科研课题的指标体系如下图所示。针对众多的候选评审课题如何进行科学选择？请使用层次分析法，确定$ C_1 $到$ C_6 $相对于总目标$ A $的权重。

對於這道例題，我們首先需要計算$ W_i $，公式為$ W_i = (\prod_{j=1}^{n}a_{ij})^{\frac{1}{n}} $，例如

隨後對其進行歸一化,得到$ {W_i}^0 $，公式為$ {W_i}^0 =\frac{W_i}{\sum_iW_i} $，例如

{% raw %} 在歸一化完成之後，我們需要進行一致性檢驗，首先需要計算$ {\lambda}{max} $，公式為$ {\lambda}{max} \approx \frac{1}{n}\sum_{i=1}^n\frac{{{(AW)}i}}{W_i} = \frac{1}{n}\sum{i=1}^n\frac{\sum_{j=1}^n{(a_{ij}W_j)}}{W_i} $,例如 {% endraw %}

$(AW)_i $的計算方式可參考下圖，類似矩陣的運算。

{% raw %} 隨後計算$ C.I. $，公式為$ C.I. = \frac{{\lambda}_{max}-n}{n-1}$，例如 {% endraw %}

查表得到$ R.I. $的值。

隨後計算$ C.R. $，公式為$ C.R. = \frac{C.I.}{R.I.}$，若$ C.R. < 0.1 $，即代表通過一致性檢驗，例如

{% raw %} 最後求取各方案的總重要度，公式為$ C_j = \sum_{i=1}^n{(b_i{C_j}^i)}$，例如 {% raw %}

結果表明，$ C_1 > C_3 > C_2 $，則方案「新建工廠」最為合適，全部計算過程如下圖。

模糊綜合評價法

模糊綜合評價法是以模糊數學為基礎，運用模糊關係合成原理，將一些辯解不清不易定量的因素定量話，從多個因素對評判事物隸屬等級狀態進行綜合評價的一種方法。

使用模糊綜合評價法建模大致可依三個步驟進行：

1. 確定因素集$ F $與評定集$ E $。因素集$ F $是評價項目或指標的集合，一般有$F = {f_i}, 1 \leq i \leq n $。評定集為評價等級的集合，一般有$ E = {e_j}, 1 \leq j \leq m $。

2. 統計、確定單因素評價的隸屬度向量，形成隸屬度矩陣$ R $。隸屬度是模糊綜合評判中最基本最重要的概念。隸屬度$ R_{ij} $指多個評價主體對某個評價對象在$ f_i $方面做出$ e_j $評定的可能性大小。

3. 確定權重向量$ W_F $，這是評價項目或指標的權重或權系數向量。此外，還可有獲得評定集的數值結果（標準滿意度矢量）$ W'_E $或權重$ W_E $。

在此使用例題進行說明。

某人在購買冰箱前，為了決定型號$ A_1 $、$ A_2 $、$ A_3 $的優先順序，讓五個家庭成員使用模糊綜合評判法進行評價。評價項目（因素）包括價格$ f_1 $、品質$ f_2 $和外觀$ f_3 $，其相應權重透過圖一的判斷矩陣獲得。評價尺度分為三個級別，例如價格分為低（0.3）、中（0.2）、高（0.1）。評判結果如圖二所示。請計算這三種冰箱的優先度並進行排序。

首先計算權重$ W_i $並進行歸一化，得到如下結果

隨後進行計算，最終結果如下圖。通過比較評價值，可確定優先順序為$ A_2 $、$ A_1 $、$ A_3 $。

那失落的三年，也如同這個帳號一樣，隨著政府的一紙公文，甚至一句口頭通知，便被直接抹去。

他們是沒有記憶的，但我不是。

我不會忘記在那三年中的記憶。

2020年初，高中一年級的寒假，在期末考落幕與正式寒假間隙，我已提前完成了所有寒假任務，然而隨著五月校園重啟，那些作業卻不再被人提及。

春節前夕，家庭聚會中，人們談論著武漢肆虐的疫情，無人會相信這場疫情會在偉大光榮正確的政府領導下走向失控。但即便懷疑其廣度，母親仍堅決阻止我乘坐高鐵前往鄰市旅遊，她深知雖本地無恙，但防範之舉不可或缺。

在微微的惶恐中，我們度過了春節，家族裡的人們分別從香港、北京等地趕回家中，一同吃了年夜飯，隨後便立刻返程——某位消息靈通的長輩指出接下來會採取「特殊的防疫措施」，若不立即返程，便有可能困在這裡。

初二，預期的化學奧賽集訓營因防疫政策戛然而止。初四，我們在線上群聊中熱烈討論著學校可能的政策調整，而校辦突然發布公告，讓我們「做好暫時居家上網課的準備」，隨後第二學期所需的教科書被批量快遞至家中，我佈置好書房的環境，準備迎接從未體驗過的「網課」。

2月2日，網課的初次登場顯得有些原始，當時學校尚未來得及全面採用釘釘或騰訊會議這樣現代化的工具，因此老師們只能生疏地使用「QQ群直播」和「QQ群語音」主持課堂。

「QQ群直播」是在群組內的直播功能，與其他平台的流媒體直播類似，觀眾的名單並不會全部展示出，且觀眾可以為「主播」（即老師）贈送禮物。面對熟悉的老師和不熟悉的平台，大部分學生都格外興奮，不斷地為老師刷著廉價的禮物——不知道老師那裡會有多少直播收益。在網課進行約一週後，「QQ群課堂」才匆匆上線，刪去了禮物功能，僅保留了直播。「QQ群語音」則類似其他平台的語音聊天功能，並無什麼特殊之處。

由於尚未引入騰訊會議，我們在網路課堂上享受著一定程度的隱私，無需開啟攝像頭，只需在被提問時才需要打開麥克風，這使得课堂上充斥著一種自由自在的氛圍。課堂期間，我們在各自的班級和年級群組裡熱烈交流，仿若實境課堂中的私下討論，我發出一條空間說說，「我們成功模仿了課堂上的私語熱絡」，獲得了半數同學的轉發。

有一日，我在瀏覽「島市老八」的短視頻時，意外被點名回答問題，那猝不及防的時刻，我迅速抓起電腦的麥克風準備應答，卻不慎遺漏了暫停手機上正在播放的視頻。于是，所有在線的同学们無意間共享了「一日三餐沒煩惱，今天就吃老八秘制小漢堡」。

我們的作業都是在線上佈置，在線上提交。政治作業因為文字中涉及簡體中文互聯網的「敏感內容」，被QQ提示「安全檢查未通過」，被禁止下載。此外，有些同學心生懶意，從他人的作業中摘錄片段，妄圖蒙混過關，孰料在匆忙中，姓名標注的疏忽暴露了真相，老師在大群通報了這些行為。

在這段時間，儘管防疫措施不斷收緊，但我們實際上依舊可以相對自由地行動。在3月10日凌晨，黎明的微光尚未消退，我們迎來了一幕天文奇觀——超級月亮。我和一位交誼深厚的，且同住於這座城市的小區的小學同學，興致勃勃地登上天台，捕捉那皎潔明亮的月亮，彷彿握住了這天體高掛深紫色天空的美好夜晚。

而翌日，我與當時的女友經過吵架後分手，但幸運的是，這場陰霾在四位摯友的調解下逐漸散去，經過一番深思熟慮與情感交流，我們最終和好。但借助每日大量互動而點亮的QQ標識，因為這一次吵架而中途熄滅。

我同這一位女友在網課期間換用過諸多情侶頭像，早期是卡通的倉鼠圖片，後面換上了對方扎朝天辮的自拍——許多老師在此時發現了我們的戀情，向我們私信並八卦起來，大都表示很看好我們兩個人（用現在的話來講，「很好嗑」），但希望我們不要因為調情過多而忽視學習。而在同女友一同觀看《異度侵入 ID: INVADED》後，我們還換上了從番劇中截圖的「洞哥」作為情侶頭像。

我們此時上課無人監管，因此我們可以盡情從事自己熱愛之事。我同女友借著流媒體平台的功能一起補全了大量番劇（據嗶哩嗶哩2020年度追番報告顯示，我超越了99.7%的用戶），其中包括《頭文字D》《JOJO的奇妙冒險》等。同時我們也沉浸於觀賞經典影片，將亞佛烈德·希區考克和史丹利·庫柏力克執導的電影悉數補完，又觀看了卓別林的諸多電影以及IMDb250榜單中的大量佳作。

在家中過於無聊，我開始嘗試自媒體運營，所經營的知乎帳號與嗶哩嗶哩的主帳號在被封禁前，分別實現了10k粉絲與23k訂閱——後因觸及「敏感話題」而陸續封禁。在嗶哩嗶哩上主要用來觀影的副帳號上傳了一些音樂相關的專欄與視頻，粉絲數漲到了1k左右，如今因為長期斷更而下滑至700上下。

四月份，學校換用騰訊會議進行網課授課，我們被要求開啟攝像頭，但這依舊沒有影響我們課堂摸魚。此時小道消息傳出我們即將返回學校，但即刻收到校辦闢謠，並宣布四月底將舉行線上的月考。這次考試中大量同學借助互聯網進行作弊，學校排名格外誇張，原本倒數的同學一躍進入前10%，引得我們私下的激烈討論。隨後，我參加了清華大學組織的THUSSAT並取得了校前十名的成績，此時我尚未因為心理問題導致學習成績下滑。

五月份，嗶哩嗶哩官方發布視頻「後浪」引起網路熱議。我在5月6日轉載網路評論「你點贊了後浪，我宣傳了入關，他拍攝了非浪，我们都是新時代的有為青年」。再想想自疫情以來，曾經得到廣泛推崇的「入關學」如今已毫無生命力，不禁感覺可笑。

開學前一日，我與當時的女友分手，這成為了我續心理問題的導火索。

隨文章附上一首江睿謹老師的短詩，

最終結果呈現

前期準備

請在GitHub、Clerk、Vercel註冊自己的帳號，此處不再演示。

配置網域

搭建個人博客的第一步，便是擁有自己的網域，如果你還沒有，可以前往Dynadot購買，在此可以使用我的優惠碼8E9O6bh6c9B8y6l以獲得五美元的優惠。

此處以在Dynadot購買nskr.blog為例。

介於Cloudflare提供了諸多優化站點訪問，保障站點安全的服務，我們在此選擇將站點託管至Cloudflare。

首先需要在Cloudflare中加入自己已購買的網域，並令其掃描現有的DNS解析。

隨後複製Cloudflare提供的DNS伺服器，回到Dynadot控制台，在DNS設定處選擇網域名稱伺服器，並填入已複製的內容。

配置伺服器

在完成有關網域的相關設定後，我們便需要進行伺服器的配置。我在此選擇的是Google Cloud，在綁定貸記卡後可以獲得300美元的贈金，有效期三個月。

根據自己的需求配置和購買，我按照個人習慣選擇了Ubuntu的系統鏡像。注意在此處需開啟「允許HTTP流量」、「允許HTTPS流量」兩個選項。

在伺服器構建完畢後，我們可以獲得這一伺服器的公網IP。我們需要複製這一IP並在Cloudflare的控制台中配置網域解析，將api.nskr.blog指向自己的伺服器。（實際上，在此處無需參照影片中的操作，同時解析@.nskr.blog，因為後面需要將其解析到Vercel的伺服器上）

在完成前面的操作後，我們通過SSH連結自己的伺服器。

為了方便管理，我們可以安裝1Panel面板或寶塔面板，二者選其一即可：

#安裝1Panel
curl -sSL https://resource.fit2cloud.com/1panel/package/quick_start.sh -o quick_start.sh && bash quick_start.sh

#安裝寶塔面板
wget -O install.sh https://download.bt.cn/install/install-ubuntu_6.0.sh && sudo bash install.sh ed8484bec

這安裝這兩個面板後都會自動安裝Docker與Docker-compose，可以通過下面的命令分別檢查：

docker -v

docker compose version

如果伺服器沒有正確輸出已安裝版本號，可以執行以下命令即可：

curl -fsSL https://get.docker.com | bash -s docker

在完成上面的步驟後，請前往伺服器控制台開放相關端口，面所需端口可以在SSH介面查詢，博客所需端口為2323與2333：

部署博客後端

首先，需要拉取配置文件：

# 創建目錄
cd && mkdir -p mx-space/core && cd $_

# 拉取 docker-compose.yml 文件
wget https://fastly.jsdelivr.net/gh/mx-space/core@master/docker-compose.yml

並在mx-space/core目錄下新建.env文件並填入以下內容（請根據自身需要進行修改）：

有關vim的使用可以自行學習，在此僅提及與部署相關的操作：

使用vim .env創建並打開.env文件，隨後輸入i切換至insert模式，在輸入完變量後按ESC鍵進入命令模式，輸入:wq保存退出。

JWT_SECRET=nishikoriyuinishikoriyui
ALLOWED_ORIGINS= nskr.blog
ENCRYPT_ENABLE=false
ENCRYPT_KEY=

• JWT_SECRET 長度不小於16個字符，不大於32個字符，用於加密JWT。
• ALLOWED_ORIGINS 為被允許的域名,如果存在多個需用英文逗號分割。
• ENCRYPT_ENABLE是否開啟加密，填寫true或false。
• ENCRYPT_KEY為加密密鑰，如不開啟加密，則無需填寫。密鑰長度需為64位且僅包含小寫字母與數字，可通過openssl rand -hex 32命令生成。

完成上面的操作後，可通過下面的命令拉取鏡像並運行後端核心容器：

docker compose up -d

完成後端部署後，我們需要配置反向代理。使用單域名的方案可以參照官方文檔 ，我在此選擇借助1Panel進行雙域名配置，寶塔面板配置方法與之類似，在此不再演示。

首先需要在1Panel安裝OpenResty程式，用於網站部署。

隨後，在「網站」選擇卡中搭建靜態網站api.nskr.blog

完成網站搭建後，需要申請SSL證書，在此不再演示。

搭建完成後，選擇「配置」，並將下面的內容粘貼至網站配置文件：

    ## 反向代理开始
    ## WebSocket
    location /socket.io {
      proxy_pass http://127.0.0.1:2333/socket.io; 
      proxy_set_header Host $host; 
      proxy_set_header X-Real-IP $remote_addr; 
      proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for; 
      proxy_set_header REMOTE-HOST $remote_addr; 
      proxy_set_header Upgrade $http_upgrade; 
      proxy_set_header Connection "upgrade"; 
      proxy_buffering off;
      proxy_http_version 1.1; 
      add_header Cache-Control no-cache; 
    }
    ## Others
    location / {
      proxy_pass http://127.0.0.1:2333; 
      proxy_set_header Host $host; 
      proxy_set_header X-Real-IP $remote_addr; 
      proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for; 
      proxy_set_header REMOTE-HOST $remote_addr; 
      add_header X-Cache $upstream_cache_status; 
    }
    ## 反向代理结束

這樣配置後，

API 地址為 https://api.nskr.blog/api/v2

GateWay 為 https://api.nskr.blog/

博客後端面板為 https://api.nskr.blog/proxy/qaqdmin

部署博客前端

在此我們使用Vercel的Serverless服務進行前端的部署，這可以有效減輕我們伺服器的壓力，從而減少在伺服器上的開支。

我們首先需要在GitHub上克隆Shiro的官方倉庫。

隨後，我們需要打開Clerk，並創建新應用，獲得公鑰和私鑰。

隨後我們需要訪問博客的後端面板，在配置雲函數。

在編輯面板中填入：

• 名稱：shiro
• 引用：theme
• 數據類型：JSON

並複製該連結的內容並填入（此處直接引用本站點的配置文件，請根據自身需要進行修改）：

在完成上述步驟之後，前往Vercel，創建新項目,在配置頁面中，点击 Environment Variables，填入以下內容：

NEXT_PUBLIC_API_URL=https://api.nskr.blog/api/v2
NEXT_PUBLIC_GATEWAY_URL=https://api.nskr.blog
NEXT_PUBLIC_CLERK_PUBLISHABLE_KEY=pk_test_•••••••••••••••••••••••••••••••••••••••••••
CLERK_SECRET_KEY=sk_test_•••••••••••••••••••••••••••••••••••••••••••

• NEXT_PUBLIC_API_URL 為API地址
• NEXT_PUBLIC_GATEWAY_URL 為Gateway地址
• NEXT_PUBLIC_CLERK_PUBLISHABLE_KEY 是Clerk獲得的公鑰
• CLERK_SECRET_KEY=sk_test_••••••••••••••••••••••••••••••••••••••••••• 是在Clerk獲得的私鑰

隨後點擊Deploy，等待部署完成

在部署完成之後，將自己的網域名稱與之綁定，並將博客連結在Cloudflare上解析到Vercel的服務器。

就此，基本的部署便已經完成。

參考資料

Mix Space官方文檔 https://mx-space.js.org/

最近幾天依舊很忙，每天都有挺多課，但從下週開始會好一些——因為 Systems Engineering 以及 Engineering Materials 兩門課程全部在這周結課，並在十五週（一個半月後）進行期末考試。

不知怎的，KFC 現在推出了「Pizza」（儘管我並不認可這個命名），Pizza Hut 則推出了漢堡。週二、週三兩天我分別嘗試了兩款新品漢堡，一款名叫「招牌秘制雞腿堡」，另一款叫做「芝士和牛至尊堡」。

::: grid {cols=2,gap=12}



:::

總的來說，作為靠「烙大餅」發家的 Pizza Hut，為這兩款漢堡烤製的麵包非常不錯，氣孔均勻，軟硬適中，還伴有一絲奶香奶甜口味。個人認為一個漢堡好不好吃，其核心就在於麵包的品質，而這兩款產品都做的很好。除此之外，「招牌秘制雞腿堡」並無其他特別之處，而「芝士和牛至尊堡」還有其他一些亮點。和牛肉餅的質量無庸置疑，而其中的配菜是讓我真正為這款漢堡打高分的原因。切片的洋蔥經過烤製，退去了自身的辛辣，保留了自身的輕脆，一口咬下去可以感受到其甘甜的汁水，減輕了牛肉堡自帶的「厚重感」。而另一款配菜青椒也十分不錯，裡面綠色的青椒片同樣經過烹飪而弱化了自身的「青草味」，而依舊具備爽脆的口感。

除了漢堡，我也嘗試了幾款小食。不得不說一塊普通的雞肩胛骨，在 Pizza Hut 手裡竟能推出這麼多款不同的小食——「蜜制醬烤香骨雞」「招牌甜辣韓式炸雞」「新奧爾良香骨雞」「香甜吮指韓式炸雞」。我這次只是嘗試了新品，但即使沒有對比，我依然要表示還是最早的「蜜制醬烤香骨雞」最對我口味。「北海道戀人白桃芝士撻」中規中矩，無法動搖我心中第一的 KFC 蛋撻的地位。

原本以為 Engineering Materials 的實驗課結束之後，之後可以輕鬆一些，但沒料到 Strength of Materials 的實驗緊接著開課。週三的早晨（又是早八）我來到學校的力學中心，完成了拉伸實驗。

實驗中，我們拉斷了低碳鋼棒和鑄鐵棒，分別用了 35kN 與 17kN 的力。

::: grid {cols=2,gap=12}

:::

至於週四，我一大早便去了 Starbucks，在那裡完成了 Strength of Materials 和 Systems Engineering 的相關作業，並且複習了一些 Engineering Materials 的作業。隨後下午六點多回到宿舍，一覺睡到了晚上十點。

晚上寫了一篇東北話編程的水文，又配置了一個雲函數，顯示我使用 Mac 的實時狀態。

忙完後計畫繼續睡覺。但不知怎的卻一直失眠，直到四五點鐘才勉強入睡。

今日瀏覽社交平台時，偶然發現一張令人會心一笑的梗圖：

透過對東北方言稍加熟悉，我們就能輕鬆將其轉換成這樣的代碼表示：

def print_fibonacci_sequence(n):
    if n < 1:
        raise ValueError("Invalid input! Please enter a positive integer.")

    fibonacci_sequence = []
    current = 0
    next_num = 1
    fibonacci_sequence.append(current + next_num)

    for _ in range(2, n+1):
        current, next_num = next_num, current + next_num
        fibonacci_sequence.append(next_num)

    return ', '.join(map(str, fibonacci_sequence))

print("Checking result for 0:")
try:
    print(print_fibonacci_sequence(0))
except ValueError as e:
    print(e)

print("Checking result for 15:")
print(print_fibonacci_sequence(15))

# 並未完全對應原文的代碼

尋根

然而，注意到梗圖中的代碼結構清晰且配有代碼高亮標記，讓人感到這并非單純的惡搞之作，于是我決定在互聯網上進一步探索，竟意外地找到了相關的實在項目：

https://github.com/zhanyong-wan/dongbei

dongbei 是啥？它是一门以东北方言词汇为基本关键字的以人为本的编程语言。

这玩意儿可是填补了世界方言编程地图上的一大片儿空地啊！ 这么说吧，谁要是看了 dongbei 程序能憋住了不笑，我敬他是纯爷们儿！

那它有啥特点咧？多了去了：

简单啊！小学文化程度就行。您能看懂春晚不？能？那就没问题。 好读啊！看着看着包您不由自主地念出声儿来。 开心啊！呃，做人嘛，最重要的是要开心。 开源啊！不但不要钱，而且不要脸 -- 随时随地欢迎东北话高手打脸指正。 总而言之，dongbei 语言具有极高的娱技比（娱乐精神-技术含量比例）。

dongbei 编程语言的开发采用了业界领先的 TDD（TreeNewBee-Driven Development） 方式。 具体地说，就是每个功能都是先把文案写好，八字没一撇牛皮就吹起来了，然后根据牛皮写测试案例，最后再实现功能让牛皮不被吹破。 这样做有两大好处：第一每个功能都是有的放矢，不值得 tree new bee 的功能一概没有。 第二确保了每个功能都有文案负责吹嘘，开发者绝对不会养在深闺无人识。

經過搜尋，我發現了一個來自該項目的示例文件：

https://github.com/zhanyong-wan/dongbei/blob/master/demo/斐波那契.dongbei

這便確定了梗圖中的代碼，來自這個名為 dongbei 的項目。

問底

根據項目的文檔指示，我將dongbei的代碼庫成功克隆至本地，執行src/dongbei.py，這樣便能親自體驗這個充滿趣味的編程語言。

然而，dongbei的功能遠不止于此。通過閱讀dongbei 语言咬文嚼字，我們可以發現這門語言已經相當完備，足以應對各種複雜的編程挑戰。

例如：

https://github.com/zhanyong-wan/dongbei/blob/master/demo/汉诺塔动唤.dongbei

https://github.com/zhanyong-wan/dongbei/blob/master/demo/活着.dongbei

融會

事實上，網絡上除了標新立異的dongbei之外，還隱藏著眾多富有創意的編程語言，它們通常被歸類為「Esolang」——Esoeric programming language，旨在嘗試奇怪的想法，難以程式設計，或作為笑話，而不是實際使用。

Cantonese

參考dongbei的模式，還有另一個創新的項目——Cantonese，它讓開發者能夠使用粵語進行編程，進一步將地方方言融入編程語言的世界，作者至今依舊在維護它。

https://github.com/StepfenShawn/Cantonese

使用該語言輸出「Hello World！」的代碼：

畀我睇下「"Hello World!"」點樣先？
收工

Ook！

Ook! 是一款極致簡潔的編程語言，其核心語法僅由三個基本構件組成，乍看之下易於學習，然而在實際應用中，Ook! 帶來了獨特的挑戰。這種語言的命令系統由 Ook.、Ook! 和 Ook? 這三個元素構成，即便代碼量可能因此增加，卻絲毫不影響創建高效能的程式碼。Ook! 的設計原則簡潔到了極致，其創建者深信甚至猩猩也能輕易掌握，因此他們堅信無需附加繁冗的程式註解。

使用該語言輸出「Hello World！」的代碼：

Ook. Ook? Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook.
Ook. Ook. Ook. Ook. Ook! Ook? Ook? Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook.
Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook? Ook! Ook! Ook? Ook! Ook? Ook.
Ook! Ook. Ook. Ook? Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook.
Ook. Ook. Ook! Ook? Ook? Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook?
Ook! Ook! Ook? Ook! Ook? Ook. Ook. Ook. Ook! Ook. Ook. Ook. Ook. Ook. Ook. Ook.
Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook! Ook. Ook! Ook. Ook. Ook. Ook. Ook.
Ook. Ook. Ook! Ook. Ook. Ook? Ook. Ook? Ook. Ook? Ook. Ook. Ook. Ook. Ook. Ook.
Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook! Ook? Ook? Ook. Ook. Ook.
Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook? Ook! Ook! Ook? Ook! Ook? Ook. Ook! Ook.
Ook. Ook? Ook. Ook? Ook. Ook? Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook.
Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook! Ook? Ook? Ook. Ook. Ook.
Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook. Ook.
Ook. Ook? Ook! Ook! Ook? Ook! Ook? Ook. Ook! Ook! Ook! Ook! Ook! Ook! Ook! Ook.
Ook? Ook. Ook? Ook. Ook? Ook. Ook? Ook. Ook! Ook. Ook. Ook. Ook. Ook. Ook. Ook.
Ook! Ook. Ook! Ook! Ook! Ook! Ook! Ook! Ook! Ook! Ook! Ook! Ook! Ook! Ook! Ook.
Ook! Ook! Ook! Ook! Ook! Ook! Ook! Ook! Ook! Ook! Ook! Ook! Ook! Ook! Ook! Ook!
Ook! Ook. Ook. Ook? Ook. Ook? Ook. Ook. Ook! Ook.

Brainfuck

與Ook!類似，Brainfuck是一種簡單的、可以用最小的編譯器來實現的、圖靈完備的程式語言。這個語言由八種運算子構成。

使用該語言輸出「Hello World！」的代碼：

++++++++++[>+++++++>++++++++++>+++>+<<<<-]
>++.>+.+++++++..+++.>++.<<+++++++++++++++.
>.+++.------.--------.>+.>.

Piet

Piet 是一種獨特的編程語言，它利用色彩的組合來表達程序的邏輯結構。該語言使用了20種亮度級別的色彩，每種色彩都承載著特定的指令功能。編譯器通過辨識色塊之間的亮度差異來解讀程序的命令。程序的布局和結構完全依賴于色塊的形狀和排列，而黑白色調的邊界則作為區分不同代碼區塊的分界線。雖然在Piet上編寫程序是可能的，但由于不同解釋器可能對色彩的解釋存在差異，這使得Piet在實際計算中并不夠精確。因此，Piet 更多地被視為一種創意工具，用于探討編程的新方法和視覺藝術的表現形式，而非作為解決實際問題的工具。它的存在挑戰了傳統編程語言的界限，讓編程過程變得更加豐富多彩和富有表現力。

Velato

Velato 是一種創新的編程語言，它允許開發者通過音樂來編寫代碼。這種語言通過分析 MIDI 文件來讀取源代碼，讓你可以在創作旋律的同時編寫程序。它根據音符的順序、間距和音高（視為編程的字母）來處理代碼。儘管 Velato 支持基本的編程結構，如表達式、數字和變量聲明，但它目前不支持函數和數組。此外，它不適用於將真正的音樂作品轉化為功能性程序。

其他語言

可以參考the esoteric programming languages wiki這一站點，了解其他的Esolang。

根據Esolang Wiki，Esolang通常分為以下幾種類別：

• Minimalism 許多神秘語言的設計目標是盡可能少的指令。例如 Brainfuck、OISC 和 Lazy K 就屬於這種類型。當這些語言實現了圖靈完備，常被稱為「圖靈泥潭」。

• New concepts 探索替代方式設計編程語言在神秘編程中很受歡迎。像 Befunge、Thue 和 Unlambda 是很好的例子。

• Weirdness 有些語言主要為了追求奇特和難以編程而創造。INTERCAL 的初衷就是要與普通語言截然不同（雖然仍有很多與傳統語言相似之處），而 Malbolge 的設計目的是極度難以使用。

• Themed 有些語言基於非計算機相關的主題。比如 var'aq 基於虛構的克林貢語。Shakespeare模仿莎士比亞戲劇風格，而 Chef 程序看起來像烹饪菜譜。dongbei也屬於此類。

• Brevity 許多神秘語言旨在設計得尽可能短小。這些被稱為「Golfing語言」，常常用於「程式碼高爾夫」，即競賽解決編程任務，盡量以最少的字符或字節完成。如 CJam、Pyth 和 GolfScript 等。

• Jokes 許多神秘語言純粹是作為笑話創建。其中一些仍然可以用於編程，如 l33t 和 Ook!，而其他如 HQ9+ 和 Bitxtreme 则不然。

• Obfuscation 有些神秘語言刻意設計得難以閱讀（而非難以書寫或理解）。Unreadable 是這方面的典型例子。

納要

dongbei是一個基於python開發的編程語言。可能有人會認為「這也太不正式了吧」。但實際上，我們需要明確，dongbei只是一個娛樂或探索為目的設計的編程語言，它並不適用於開發嚴肅的應用程序。在開發需要穩定性、容錯能力和功能性的項目時，幾乎所有人都只會去選擇python，C++，JavaScript等成熟的、穩定的語言。

鑄鐵是人類使用最早的金屬材料之一，具有生產方法簡便，成本低廉，性能優良等特點，至今仍使用廣泛。缺點：硬脆，不易加工，只能鑄造。

典型零件：機床床身、汽缸體和缸套、凸輪軸、曲軸。

鑄鐵的類別：

1. 白口鑄鐵：碳主要以滲碳體（$Fe_3C$）形式存在的鑄鐵。

   • 硬而脆，很少用做零件，作原料用（生鐵）。

2. 灰口鑄鐵：碳主要以石墨（G）形式存在的鑄鐵。

3. 麻口鑄鐵：介於白口鑄鐵和灰口鑄鐵之間鑄鐵。

鑄鐵的石墨化

$Fe-C $和 $Fe-Fe_3C$ 雙重相圖

鑄鐵的石墨化過程

冷卻（結晶）時的石墨化過程

1. 從液相中析出石墨

   • $L$ → $G_I$
   • $L$ → $\gamma + G$（共晶反應）

2. 從奧氏體中析出石墨

   • $\gamma$ → $G_{II}$

3. 由共析反應生成石墨

   • $\gamma$ → $\gamma + G$

加熱時的石墨化過程

$Fe_3C$ → $3Fe + C$

鑄鐵石墨化過程的兩個階段：

• 第一階段石墨化：發生在 P'S'K'線以上的石墨化過程。
• 第二階段石墨化：發生在 P'S'K'線以下的石墨化過程。

鑄鐵組織的形成條件

按兩個階段石墨化進行的程度不同，鑄鐵具有不同的組織。

常用鑄鐵的牌號和性能特點

一般來說，白口鑄鐵、麻口鑄鐵使用不多，灰口鑄鐵應用較多一些。

灰口鑄鐵的類別

按石墨的形態，灰口鑄鐵分為四類。

• 灰鑄鐵：石墨呈片狀

• 球墨鑄鐵：石墨呈球狀

• 蠕墨鑄鐵：石墨呈蠕蟲狀

• 可鍛鑄鐵：石墨呈團絮狀

石墨狀態主要受鑄鐵的化學成分及工藝過程影響，片狀、球狀和蠕蟲狀在鑄造後即可形成，團絮狀由白口鐵退火得來。

:::grid{cols=2,gap=12，type=images}

:::

灰口鑄鐵的組織

組織特徵：鋼基體上分佈著不同形態的石墨。

• F＋G 鋼基體（鐵素體）和石墨

• F＋P＋G 鋼基體（鐵素體＋珠光體）和石墨

• P＋G 鋼基體（珠光體）和石墨

灰口鑄鐵的性能

主要取決於石墨的形態、大小、數量。

灰鑄鐵

最便宜，應用最廣泛，80% 以上。

灰鑄鐵的牌號

HT×××。“HT”表示“灰鐵”，“×××”表示最低抗拉強度值。

• 例：H200
  • HT 為“灰鐵”的漢語拼音字頭
  • 200 表示該灰鑄鐵的抗拉強度值不小於 200MPa

灰鑄鐵的性能特點

1. 力學性能低

   • 片狀石墨對鋼基體的分割作用大，尖端形成應力集中。

2. 耐磨性和減振性好

   • 石墨的存在，有利於潤滑和貯油、吸收振動能量，減震性優於碳鋼。

3. 工藝性能好

   • 熔點低，液態流動性好，易於澆注，特別是結構複雜鑄件；切削時易斷屑（切削性能優於鋼）。

孕育處理

加孕育劑（變質劑）：矽鐵合金、矽鈣合金，結晶核心多，石墨片尺寸更小更均勻。

應用

機床床身、底座、發動機缸體等。

球墨鑄鐵

改變石墨形態來提高力學性能。灌鑄前加球化劑和孕育劑，經過球化處理和孕育處理得到。

球化劑：鎂、稀土、稀土鎂。

球墨鑄鐵的牌號

QT×××-××。“QT”表示“球鐵”，“×××”表示最低抗拉強度值，“××”表示最低延伸率。

• 例：QT500-05
  • QT 為“球鐵”的漢語拼音字頭
  • 500 表示該球墨鑄鐵的抗拉強度不低於 500MPa
  • 05 表示該球墨鑄鐵的延伸率不低於 5%

球墨鑄鐵的性能特點（與灰鑄鐵相比）

1. 力學性能高

   • 球墨鑄鐵抗拉強度和彎曲疲勞極限高、塑性和韌性好。
   • 球狀石墨對鋼基體的分割作用減低至最小，鋼基體的連續性好，應力集中減弱。

2. 減振性差

   • 減震性不如灰口鐵。

3. 鑄造性差

應用

一定條件下可替代鑄鋼、鍛鋼，用以受力複雜、負荷較大和要求耐磨的鑄件，如內燃機曲軸、凸輪軸閥門、汽車後橋殼等。

蠕墨鑄鐵

鐵液經變質處理和和孕育處理得到。

變質元素（蠕化劑）：稀土矽鐵鎂合金、稀土矽鐵鎂合金、稀土鈣鐵合金等。

蠕墨鑄鐵的牌號

RuT×××。“RuT”表示“蠕鐵”，“×××”表示最低抗拉強度值。

球墨鑄鐵的性能特點

1. 力學性能介於灰鑄鐵和球墨鑄鐵之間

   • 強度和韌性高於灰鑄鐵，不如球墨鑄鐵；耐磨性較好；減振能力優於球墨鑄鐵。
   • 蠕蟲狀石墨的頭部鈍而圓，對基體分割作用較灰鑄鐵有明顯降低。

2. 工藝性能介於球墨鑄鐵和灰鑄鐵之間

   • 鑄造性能優於球墨鑄鐵，與灰鑄鐵接近。

3. 導熱性能接近灰鑄鐵

   • 高溫強度、熱疲勞性能大大優於灰鑄鐵。
   • 適於製造承受交變熱負荷的零件。

可鍛鑄鐵

可鍛鑄鐵的獲得：

1. 獲得白口鑄鐵；

2. 白口鑄鐵石墨化。

   • 可鍛鑄鐵在加熱時通過 Fe3C 的石墨化而獲得。

   • 灰鑄鐵、球墨鑄鐵、蠕墨鑄鐵在冷卻時通過析出石墨而獲得。

可鍛鑄鐵的牌號

1. KTH×××-××。“KTH”表示“可鐵黑”，稱黑心可鍛鑄鐵，以鐵素體為基體。
2. KTZ×××-××。“KTZ”表示“可鐵珠”，稱珠光體可鍛鑄鐵，以珠光體為基體。

可鍛鑄鐵的性能特點

性能介於灰鑄鐵與球墨鑄鐵之間，耐蝕性較好，但生產效率低。

鑄鐵的熱處理

鑄鐵熱處理的目的：

1.改變鋼基體的組織，改善鑄鐵性能；

2.消除鑄件應力。

特別提示：

1. 熱處理只能改變鋼基體的組織，不能改變石墨的形態及分佈

   • 片狀或球狀的石墨不會因熱處理而變成其他的形狀。

   • 石墨的尺寸不會因熱處理而增大或減小。

   • 石墨的分佈不會因熱處理而改變。

2. 灰鑄鐵不適合採用強化型（如淬火）熱處理

   • 灰鑄鐵中石墨呈片狀，對基體的分割作用很大，即使進行強化型熱處理也難以獲得明顯的效果。

3. 球墨鑄鐵適合採用與鋼相同的各種熱處理

   • 球墨鑄鐵中石墨呈球狀，對基體的分割作用小，因此通過熱處理可顯著改善其力學性能。

時效處理

目的：釋放鑄件應力。

工藝：

• 人工時效：將鑄件加熱到 500～560°C 保溫後隨爐冷卻。

• 自然時效：將鑄件放在室外 6～18 個月，使應力自然釋放。

人工時效是目前生產中最常用的方法。

消除白口退火

目的：消除白口組織。

鑄件表面或薄壁處在鑄造過程中會因冷卻速度過快而出現白口組織，其硬度很高，給切削加工帶來困難，必須予以消除。

工藝：

• 加熱溫度：880～900°C。
• 保溫時間：1～2h。
• 冷卻方式：保溫後緩慢冷卻至 400～500°C 後出爐空冷。

表面熱處理

目的：提高鑄件的表面硬度、耐磨性和耐蝕性。

工藝：

• 感應加熱表面淬火

• 激光加熱表面淬火

• 滲氮、滲金屬

球墨鑄鐵的熱處理

退火

目的：提高球墨鑄鐵件的韌性。

工藝：

1. 880～900°C 加熱保溫後，爐冷至 600°C 出爐空冷。 - 適應的鑄件：含有白口組織的鑄件。 - 退火後的組織：$F＋G_球$

2. 700～760°C 加熱保溫後，爐冷至 600°C 出爐空冷。

   • 適應的鑄件：組織為$F＋P＋G_球$的鑄件。
   • 退火後的組織：$F＋G_球$

正火

目的：將基體轉換為細珠光體組織，以提高強度、硬度和耐磨性。

工藝：850～900°C 加熱保溫後出爐空冷。

• 適應的鑄件：組織為 F＋P＋G 球的鑄件。
• 正火後的組織：$P＋G_球$

淬火及回火

目的：提高球墨鑄鐵的力學性能。

1. 淬火（860～900°C）→ 低溫回火（250～350°C） - 組織：$M_回＋A'＋G_球$。可用於製造軸承。
2. 淬火（860～900°C）→ 中溫回火（500～600°C） - 組織：$S_回＋G_球$。可用於製造軸類零件。

等温淬火

目的：獲得$B_下$，使球墨鑄鐵具有良好的綜合力學性能。

工藝：830～870°C 加熱保溫後投入 280～350°C 的熔鹽中保持。

組織：$B_下＋A'＋G_球$

在課上摸魚搭了兩個新的小玩具。一個是文件轉換，可以幫我快捷的轉換文件格式。另一個是 freegpt，從 chatgpt 的網頁端提取 api-key 從而免費使用 gpt-3.5，這個因為 openai 的用量/請求次數限制，就不打算設為公開了。

在將DNS遷移到Cloudflare之後，某些頁面的JavaScript內容渲染總是出現錯誤。今天終於解決了這一問題。Cloudflare為優化網頁訪問，有一個名為Rocket Loader™的功能，作用是「改善含有 JavaScript 的網頁繪製時間」，但在實際中會導致渲染錯誤。在關掉這一功能後，一切恢復正常。

今天格外的喜歡聽草東沒有派對的《滔滔》，台語很有意思，但我至今還沒學會多少。

https://www.youtube.com/watch?v=GavvkJ-Svrg

晚飯依舊是自己煮的麵，不過不是方便麵，而是一種類似掛麵的。味道還可以，價格也很好。除了麵之外，還自己做了一杯奶茶，味道也還可以，我不愧是廚藝小天才。

吃完晚飯後，斷斷續續完成了 Mechanisms & Machine Theory 的作業，十分的折磨，因為解題不僅需要帶入公式，還需要精確的作圖。在做到一半時實在忍不下去而選擇摸魚，下樓轉了一圈，順便買了一包菸——以前買菸我總是去一公里外的熙街裡的一家便利店，今天不想走那麼遠，於是就去了校門對面的超市碰碰運氣，才意外發現原來這家超市也賣香菸，看來以後會便利很多了。

明天要從早上八點鐘上課上到下午四點，後面如果有時間的話準備動筆寫一下對癸卯年國慶假期的回憶的第五篇。這一篇因為會有些色情敏感（x）的內容，於是依舊是加密的。

除了昨天提到的幾部網劇，最近也在看《漫長的季節》和《白夜追凶》，不得不說這幾部劇改變了我對中國產網劇的印象

重慶的春天向來是這樣，有時烈日當空，氣溫直升至三十多攝氏度，有時又連續幾日陰雨，寒風凜人，見不到一點陽光。在這樣的情緒影響下，我的心情總是很不好，希望春天能儘快過去，夏天儘管炎熱，但見了日光，大概我的情緒也會積極起來。

好想要和人貼貼。

週四的時候觀察了常用材料的顯微圖像，拍下了一張最特別的球碳石墨。

週五都上了很多課，十分疲憊，在空閒之時完成了監控面板的部署並介入了Telegram的Bot，以後我所部署的各項服務出現問題都可以及時發現，便於維護。

此外，還搭建了屬於自己的gpt-academic，便利個人的學術寫作，使用自己的oneapi服務接入了很多主流大模型。其中內置了LaTeX和TTS服務，功能很強，且相比於在其他公共平台進行部署，安全性更高。

週六休息到了中午才起床，下午去實驗室做了實驗，這個實驗因為很特殊，原本我們分了十五組，但最終能成功讀取數據的只有四個組（其中包括我自己的組）。

隨後去了商場買了幾件新衣服，還嘗了一下瑞幸的新品。

今天學習了半天，晚上在KTV開心地唱了三個小時。

晚上折騰服務器時不小心出了故障，不過萬幸我製作了快照，對服務器立刻進行了恢復。

最近看了幾部中國製作的有趣的網絡短劇《無證之罪》《隱秘的角落》和《沈默的真相》，很有意思。

在工業用鋼中，除 Fe 和 C 元素外，還含有其他元素，包括雜質元素和合金元素。

雜質元素：泛指鋼冶煉後殘留在鋼中的元素。

合金元素：指為提高鋼的性能特意添加到鋼中的元素。

碳鋼是$w_C $為 0.0218%～2.11% 的鐵碳合金，具有較好的性能，產量大、價格低，應用十分廣泛，但淬透性、回火抗力差、屈強比低。

合金鋼是在碳鋼的基礦上添加合金元素的鐵基合金，總體性能優於碳鋼，但加工困難、價格昂貴。選材時，如果碳鋼能夠滿足要求，一般不使用合金鋼。

鋼的分類和編號

鋼的分類

按化學成分分類

按用途分類

按金相組織分類

按質量分類

注： 高級優質鋼在其編號後加註“A” 特級優質鋼在其編號後加註“E”

鋼的編號

1. 普通碳素結構鋼——用“符號 Q＋數字”表示

• 例：Q235
• Q 為屈服點的漢語拼音字頭
• 235 表示該鋼的屈服強度值不小於 235MPa

2. 優質碳素結構鋼——用“2 數字”表示

• 例：45
• 45 表示$\omega_C$的萬分之幾，即該鋼的平均$\omega_C$為 0.45%

3. 合金結構鋼——用“2 數字＋元素符號＋數字”表示

• 例：36Mn2Si
• 36 表示$\omega_C$的萬分之幾，即該鋼的平均$\omega_C$為 0.36%
• Mn2 表示該鋼含有錳元素，平均$\omega_{Mn}$為 2%(百分之幾)
• Si 表示該鋼含有矽元素，平均$\omega_{Si}$小於 1.5%

4. 碳素工具鋼——用“T＋數字”表示

• 例：T12A
• T 為“碳素工具鋼”的漢語拼音字頭
• 12 表示$\omega_C$的千分之幾，即該鋼的平均$\omega_C$為 1.2%
• A 表示該鋼為高級優質鋼

5. 合金工具鋼——用“1 數字＋元素符號＋數字”表示

• 例：9Mn2V
• 9 表示$\omega_C$的千分之幾（當$\omega_C$>1.0%,則不予標出）
• Mn2 表示該鋼含有錳元素，平均$\omega_{Mn}$為 2%(百分之幾)
• V 表示該鋼含有釩元素，平均$\omega_{V}$小於 1.5%

6. 滾動軸承鋼——用“G＋數字”表示

• 例：GCr15
• G 為“滾動軸承鋼”的漢語拼音字頭
• Cr15 表示該鋼含有鉻元素，平均$\omega_{Cr}$為 1.5%(Cr 千分之幾，其餘百分之幾)

7. 不銹鋼——用“數字＋元素符號＋數字”表示

• 例：1Cr18Ni9Ti
• 1 表示平均$\omega_C$的千分之幾
• Cr18 表示該鋼含有鉻元素，平均$\omega_{Cr}$為 18%
• Ni9 表示該鋼含有鎳元素，平均$\omega_{Ni}$為 9%
• Ti 表示該鋼含有鈦元素，平均$\omega_{Ti}$小於 1.5%

8. 高速工具鋼——用“元素符號＋數字”表示，不標出含碳量

• 例：W18Cr4V
• W18 表示該鋼含有鎢元素，平均$\omega_{W}$為 18%
• Cr4 表示該鋼含有鉻元素，平均$\omega_{Cr}$為 4%
• V 表示該鋼含有釩元素，平均$\omega_{V}$小於 1.5%

鋼中的雜質及合金元素

雜質元素對鋼性能的影響

1. 錳（Mn）

• Mn 能溶入鐵素體中起固溶強化作用，提高鋼的硬度和強度。

2. 矽（Si）

• Si 能溶入鐵素體中起固溶強化作用，提高鋼的硬度和強度。

3. 硫（S）

• S 通常是有害元素，會引起鋼的“熱脆”。

4. 磷（P）

• P 通常是有害元素，會引起鋼的“冷脆”。

5. 氮（N）

• 過量 N 會使鋼的塑性、韌性顯著降低，出現“時效脆化”。

6. 氧（O）

• 氧通常是有害元素，在鋼中成為非金屬夾雜物而降低鋼的性能。

7. 氫（H）

• 氫是鋼中的有害元素，產生“氫脆”和造成“白點”（微裂紋）。

合金元素在鋼中的作用

合金元素在鋼中的兩種存在形式：溶入鐵素體和形成碳化物。

合金元素對鋼中基本相的影響

1. 溶入鐵素體——非碳化物形成元素

• Ni、Si、Al、Co、Cu、N、B 等非碳化物形成元素主要溶入鐵素體中，起固溶強化作用，提高鋼的硬度和強度。Ni 還能提高鋼的韌性。

2. 形成碳化物——碳化物形成元素

• Ti、Zr、V、Nb、W、Mo、Cr、Mn 等在碳化物形成元素在鋼中與 C 作用形成碳化物。

• 鋼中的碳化物通常能夠提高鋼的強度、硬度和耐磨性。

碳化物形成元素與碳的親和力的大小：Ti, Zr, V, Nb, W, Mo, Cr, Mn, ＞Fe 依次減弱

合金元素對鐵碳相圖的影響

鋼中加入合金元素後，Fe－$Fe_3C $相圖將發生變化，這些元素或者擴大奧氏體區域，或者縮小奧氏體區域。

1. 擴大奧氏體區的元素

• Ni、Mn、C、N 等元素擴大奧氏體區域，使$A_1$和$A_3$溫度降低、S 點和 E 點向左下方移動。

2. 縮小奧氏體區的元素

• Cr、Mo、Si、W 等元素縮小奧氏體區域，使$A_1$和$A_3$溫度升高、S 點和 E 點向左上方移動。

合金元素對鋼的熱處理的影響

1. 對奧氏體化的影響

• 除 Ni、Co 外，大多數合金元素都延緩鋼的奧氏體化過程。

• 它們阻礙 C 和 Me 的擴散，因此合金鋼的 A 化溫度較高、時間較長。

2. 對奧氏體晶粒度的影響

• 除 Mn 外，幾乎所有合金元素都細化晶粒。

• 尤其以碳化物形成元素的細化晶粒效果最顯著，它們形成的碳化物分布在 A 中時，能夠阻礙晶界的遷移，從而阻止晶粒長大。

3. 對鋼的淬透性的影響

• 除 Co 外，大多數合金元素只要溶入奧氏體中，都提高淬透性。

4. 對鋼的回火穩定性的影響

• 回火穩定性：淬火鋼在回火過程中抵抗硬度下降的能力。又稱回火抗力。

• 大多數合金元素因為能夠阻礙 M 的分解，從而提高回火抗力。

鋼結構

工程組件用鋼

工程組件用鋼是國防、化工、石油、電站、車輛、船舶、橋梁、建築等國民經濟部門製造各種重要組件的金屬材料。

工程組件用鋼的性能要求

1. 高的彈性極限和高的剛度，低的屈強比。
2. 良好的焊接性和冷變形性。
3. 良好的耐大氣、水腐蝕性。

工程組件用鋼的成分特點

低含碳量：$\omega_C$＜0.25%。

含微量合金元素：V、Ti、Nb、Zr、Ca、RE（稀土元素）等。

工程組件用鋼分類

普通碳素結構鋼

• 化學成分特點：低碳：0.10～0.20%
• 典型鋼號：Q235
• 熱處理特點：一般不進行熱處理強化，只保證力學性能和工藝性能
• 使用狀態：熱軋態，組織：F＋P(少量)
• 應用：建築鋼筋

低合金高強度鋼（普通低合金結構鋼）

• 化學成分特點：低碳：0.10～0.20%，錳含量：1.2～1.6%
• 典型鋼號：16Mn
  • 相當於是在 Q235 的基礦上多添加了 0.6～0.8% 的 Mn
• 熱處理特點——一般不進行熱處理強化，只保證力學性能和工藝性能
• 使用狀態：熱軋態或正火態，組織：F＋P(少量)。
• 應用：橋梁鋼構、船用鋼板等。

機器零件用鋼

機器零件用鋼是製造各種機器零件，如齒輪、軸、彈簧、軸承等的金屬材料。

機器零件用鋼的性能要求

1. 高的強度和塑性、韌性。
2. 良好的疲勞強度、耐磨性。
3. 良好的切削加工性能和熱處理性能。

機器零件用鋼的成分特點

除 C 元素外，通常含有能夠提高淬透性和細化晶粒的合金元素：Cr、Ni、Mn、Si、Mo、Ti、V、W 等。

機器零件用鋼分類

滲碳鋼

• 化學成分特點：低碳：0.1～0.25%；合金元素：Cr、Mn、Ni、Mo、Ti、W、V、B 等。
  • 滲碳後，表層為 1.0% 左右的高碳成分，保證高硬度、高耐磨性；心部保持原始成分，熱處理後具有良好的韌性和足夠的強度。
• 典型鋼號：20CrMnTi
  • 在 20 號鋼的基礎上，通過添加合金元素 Cr、Mn、Ti 而形成
• 熱處理特點：滲碳（$920～930^\circ C$）→ 淬火（$830^\circ C$）→ 低溫回火（$200^\circ C$，2h）。
• 組織和性能
  • 表層：高碳$M_回＋k＋A'$，硬度：HRC56～62。
  • 心部：低碳$M_回$（淬透時），硬度：HRC30～45。
• 應用：重載齒輪、凸輪、活塞銷等。

調質鋼

• 化學成分特點：中碳：0.30～0.50%；合金元素：Cr、Ni、Mn、Mo、W、V、Ti、B 等。熱處理後可滿足綜合力學性能的要求。
• 典型鋼號：40Cr
  • 在 40 號鋼的基礎上，通過添加合金元素 Cr 而形成
• 熱處理特點：調質處理：淬火（$830^\circ C$）→ 高溫回火（$525^\circ C$）。
• 組織：$S_回$，硬度：HRC25～35。
• 應用：各種軸類零件、汽車後橋半軸、連桿、螺栓等。

彈簧鋼

• 化學成分特點：中、高碳：0.45～0.70%；合金元素：Si、Mn、Cr、V 等。
• 典型鋼號：
  • 熱成形彈簧（彈簧用鋼的尺寸較大）：60Si2Mn（在 60 號鋼的基礎上，通過添加合金元素 Si、Mn 等元素而形成）
  • 冷成形彈簧（彈簧用鋼絲的直徑一般小於 10mm）：50CrVA。
• 熱處理特點
  • 熱成形彈簧：淬火（$870^\circ C$）→ 中溫回火（$460^\circ C$）。
  • 冷成形彈簧：卷繞→ 去應力退火（$250^\circ C$）。
• 組織：$T_回$，硬度：HRC35～45。
• 應用：
  • 60Si2Mn —— 汽車減振板彈簧、懸擺彈簧。
  • 50CrVA —— 汽車發動機氣閥彈簧。

滾動軸承鋼

• 化學成分特點：高碳：0.95～1.15%；合金元素：Cr（主要）、Si、Mn 等。
• 典型鋼號：GCr15
• 熱處理特點：
  • 預備熱處理：球化退火，以降低硬度，改善切削加工性，並獲得 P 粒。
  • 最終熱處理：淬火（840°C）→ 低溫回火（150～160°C）。
• 組織：$M_回＋k＋A'$硬度：＞HRC61。
• 應用：滾動軸承的滾珠、軸承套等。

工具鋼

刃具鋼

刃具鋼是製造各種車刀、銑刀、鑽頭等切削工具的金屬材料。

刃具鋼的性能要求

1. 高的硬度和耐磨性。HRC＞60。

2. 高的熱硬性（紅硬性）。

   • 熱硬性：刃具在高溫下保持高硬度的能力。

3. 足夠的塑性和韌性。

刃具鋼的成分特點

1. 高碳：以保證高的硬度，並形成足夠多的碳化物以提高耐磨性。

2. 合金元素：

• 提高淬透性的元素 Cr、Ni、Mn、Si 等；
• 提高熱硬性的元素 W、Mo、V、Co 等。

刃具鋼分類

碳素工具鋼

• 化學成分特點

  • 高碳：0.65～1.35%。
    • 高的含碳量保證高硬度及形成足夠的碳化物，提高耐磨性。

• 典型鋼號

  • T7、T8、T10、T12。

• 熱處理特點：

  • 預備熱處理：
    • 球化退火（T8、T10、T12）；
    • 完全退火（T7）。
  • 最終熱處理：淬火（780°C）→ 低溫回火（180°C）（除 T7 外，溫度$A_{c3}$以上）。

• 組織和性能：

  • 組織：$M_回＋k＋A'$
  • 硬度：＞HRC60。

• 應用：

  • T7、T8 —— 木工工具、刃具（切削軟金屬）及錘子、沖頭等沖擊負荷工具。
  • T9~T11 —— 手用鋸條、要求不高的刀具、鑽頭等。
  • T12、T13 —— 耐磨性最高，韌性最低，銼刀、量具等。

低合金工具鋼

• 化學成分特點

  • 高碳：0.75～1.5%；合金元素：Cr、Mn、Si、W、V 等。
    • 高的含碳量保證高硬度及形成足夠的碳化物，提高耐磨性和熱硬性。

• 典型鋼號

  • 9SiCr

• 熱處理特點

  • 預備熱處理：球化退火
  • 最終熱處理：淬火（830～860°C）→ 低溫回火（180～200°C）。

• 組織和性能

  • 組織：$M_回＋k＋A'$
  • 硬度：HRC61～63。

• 應用

  • 丝錐、板牙、鑽頭、鉸刀等。

高速鋼

在高速切削過程中，刀具的刃部溫度可達 600℃以上，低合金鋼刃具已不能滿足這種要求。

• 化學成分特點

  • 高碳：0.70～1.6%；合金元素：W、Mo、Cr、V、Co 等。
    • 高的含碳量形成足夠的碳化物，保證淬硬性；W 提高熱硬性。

• 典型鋼號

  • W18Cr4V

• 熱處理特點

  • 鍛造（使大塊的碳化物破碎）→ 等溫退火（等同於球化退火）→ 淬火（1280°C）→ 回火（560°C，3 次）（消除 A'，得 M，消除二次應力）。

• 組織和性能

  • 組織：$M_回＋k＋A'$
  • 硬度：＞HRC62。

• 應用

  • 高速車刀、拉刀、鑽頭等。

模具鋼

模具鋼是製造各種衝模、擠壓模、鍛模等模具的金屬材料。

根據模具的工作條件，模具鋼分為冷作模具鋼和熱作模具鋼兩類。

模具鋼的性能要求

保持尺寸精度（或尺寸穩定性）是量具最基本的性能要求。

1. 高的硬度和耐磨性。HRC58～60。

2. 高的組織穩定性。儘量減少殘余奧氏體量。

3. 良好的耐腐蝕性。

模具鋼的成分特點

1. 冷作模具鋼

• 高碳：以保證高的硬度，並形成足夠多的碳化物以提高耐磨性。
• 合金元素：Cr、Mo、V、W 等，提高淬透性、細化晶粒。

2. 熱作模具鋼

• 中碳：以保證良好的綜合力學性能。
• 合金元素：Cr、Ni、Mn、W、Mo、V 等。
  • Cr、Ni、Mn、Mo 提高淬透性；
  • Cr、Mo 提高回火抗力；
  • Cr、W、Mo 縮小$\gamma$區，提高$A_1$溫度，提高抗冷熱蠕變性能。

模具鋼分類

冷作模具鋼（在冷態下變形的模具）

• 化學成分特點

  • 高碳：1.4～2.3%；合金元素：Cr、Mo、V、W 等。
    • 高的含碳量保證形成足夠的碳化物，提高硬度及耐磨性。

• 典型鋼號

  • 碳素工具鋼、低合金工具鋼、高碳高鉻的 Cr12 型模具鋼。
    • Cr12MoV（Cr12 型模具鋼）

• 熱處理特點

  • 預備熱處理：球化退火（850～870°C）
  • 最終熱處理：淬火（980～1030°C）→ 低溫回火（160～180°C）。

• 組織和性能

  • 組織：$M回＋k＋A'_{ (少量)}$
  • 硬度：HRC61～63。

• 應用

• 冷沖模、冷擠壓模、拉絲模等。

量具鋼

量具鋼是製造計量工件尺寸的工具的金屬材料。

量具鋼的性能要求

保持尺寸精度（或尺寸穩定性）是量具最基本的性能要求。

1. 高的硬度和耐磨性。HRC58～64。

2. 高的組織穩定性。盡量減少殘余奧氏體量。

3. 良好的耐腐蝕性。

量具鋼的成分特點

1. 高碳：以保證高的硬度，並形成足夠多的碳化物以提高耐磨性。

2. 合金元素：Cr、Mn、W 等，提高淬透性、耐磨性。

量具鋼分類

量具鋼

• 化學成分特點

  • 高碳：0.9～1.5%；合金元素：Cr、Mn、W 等。
    • 高的含碳量保證形成足夠的合金碳化物提高硬度及耐磨性。

• 典型鋼號

  • 低合金工具鋼：CrWMn、GCr15
  • 碳素工具鋼：T10、T12

• 熱處理特點

  • 球化退火→淬火→冷處理 (淬火後繼續冷至－70～－80°C，減少 A')→長時間低溫回火（提高鋼的組織穩定性）→人工時效 (低溫回火後在 120～130°C 長時間保溫，進一步穩定尺寸)

• 組織和性能

  • 組織：$M回＋k＋A'_{ (少量)}$
  • 硬度：HRC58～64。

• 應用

  • 游標卡尺、千分尺、直尺、塞尺、塊規等。

特殊鋼

不鏽鋼

不鏽鋼是能抵抗大氣或酸等化學介質腐蝕的鋼種。

不鏽鋼又稱不銹耐酸鋼。

金屬腐蝕的形式：

1. 化學腐蝕

   • 是金屬與介質直接發生化學作用的結果，如氧化。

2. 電化學腐蝕

   • 是金屬在電解質中由於形成原電池而引起的腐蝕。

電化學腐蝕是金屬腐蝕的主要形式。

合金中的原電池

在合金中，由於組成合金的相不同，而不同相的電極電位有高有低，因此就會形成無數個原電池。

例如：鋼中的珠光體，由 F 和$Fe_3C$組成，F 的電極電位低，為陽極，$Fe_3C$的電極電位高，為陰極，如果鋼的表面覆蓋了酸性介質（如酸雨），則作為陽極的 F（鋼的基體）就會被腐蝕掉，從而造成鋼的腐蝕。

防止金屬電化學腐蝕的措施

• 室溫下獲得單相組織
  • 例如：加入＞24%Ni，室溫下可獲得單相奧氏體組織。
• 提高金屬基體的電極電位
  • 例如：加入＞13%Cr，F 的電極電位從－0.56V 跃增至 0.2V。
• 金屬表面形成鈍化膜
  • 例如：加入 Cr、Al、Si 可在金屬的表面形成致密的氧化膜，將金屬與介質分隔開。

不銹鋼分類

馬氏體不銹鋼

• 化學成分特點

  • 碳含量：0.1～0.45%；鉻含量：12～14%（提高 F）

• 典型鋼號

  • Cr13 型：1Cr13、2Cr13、3Cr13、4Cr13
  • 隨 C%的增加，強度、硬度增大，但耐腐蝕性降低。

• 熱處理特點及組織

  • 預備熱處理：完全退火或高溫回火。
  • 最終熱處理：淬火（1000～1100°C）→ 回火。
  • 對 1Cr13、2Cr13：高溫回火（600～700°C），組織：$S_回$。
  • 對 3Cr13、4Cr13：低溫回火（200～300°C），組織：$M_回$。

• 應用

  • 1Cr13、2Cr13：汽輪機葉片、水壓機閥、鍋爐管附件等耐蝕結構鋼。
  • 3Cr13、4Cr13：醫用手術刀、耐蝕彈簧等耐蝕工具鋼。