Add LCS vs post-processing comparison to READMEs

README.md

@@ -136,6 +136,21 @@ Enter comma-separated hex colors (e.g., `#FF0000,#00FF00,#0000FF`). Each color a
 - **start_step / end_step**: Step range for intervention. Paper optimal: 8–10 of 50 steps.
 - **mask**: Optional. Bilinearly downsampled to patch grid for localized control.
 
+## LCS vs Post-Processing
+
+LCS operates **during** diffusion sampling, not after — this is the key difference from traditional color grading.
+
+| | Post-Processing | LCS |
+|---|---|---|
+| **When** | After VAE decode, in pixel space | During sampling, in latent space |
+| **Mechanism** | Color filter on the final image | Modifies 3D color subspace mid-generation |
+| **Model awareness** | None — structure is already locked | Model adapts to color shifts in subsequent steps |
+| **Result** | Colors can look "painted on" — shadows/skin tones may shift unnaturally | Colors look like the model intended them — content harmonizes naturally |
+
+Example: for a warm orange sunset, post-processing tints everything orange (muddying shadows), while LCS nudges colors early in sampling so the model generates clouds, lighting, and reflections that are *coherent* with warm tones.
+
+The paper's core insight: color and structure are **orthogonal** in the latent patch space, so you can steer one without disturbing the other — impossible in pixel space where they are entangled.
+
 ## How It Works
 
 1. **Project**: Convert denoised prediction to 64D patch space, project onto 3D LCS basis

README_zh.md

@@ -135,6 +135,21 @@ LCS Load Data → LCS Color Batch → KSampler
 - **start_step / end_step**：干预步骤范围。论文最优：50 步中的第 8–10 步。
 - **mask**：可选。双线性下采样至 patch 网格分辨率，用于局部控制。
 
+## LCS 与后处理的区别
+
+LCS 在扩散采样**过程中**操作，而非生成后——这是与传统调色的核心区别。
+
+| | 后处理 | LCS |
+|---|---|---|
+| **时机** | VAE 解码后，在像素空间 | 采样过程中，在潜在空间 |
+| **机制** | 对最终图像施加颜色滤镜 | 在生成中途修改 3D 颜色子空间 |
+| **模型感知** | 无——结构已经定型 | 模型在后续步骤中自适应颜色偏移 |
+| **效果** | 颜色可能看起来像"涂上去的"——阴影/肤色容易不自然 | 颜色像模型本来就打算这样生成——内容自然协调 |
+
+例子：想要暖橙色调的日落，后处理会给整张图叠橙色（阴影变脏），而 LCS 在采样早期推动颜色，模型生成的云层、光照、反射都与暖色调**内在一致**。
+
+论文核心发现：颜色和结构在潜在 patch 空间中是**正交**的，可以单独控制颜色而不干扰结构——在像素空间中无法做到，因为颜色和结构是纠缠在一起的。
+
 ## 工作原理
 
 1. **投影**：将去噪预测转换到 64D patch 空间，投影到 3D LCS 基底