MOSS | H3 | KREA2 | QW21 | LTX25 | ideo4 Workflows

This workflow uses LCS (Latent Color Subspace) nodes for in-sampling color andsharpness control.ComfyUI implementation: facok/ComfyUI-LCS (MIT License).Official code: ExplainableML/LCS.Research: Mateusz Pach, Jessica Bader, Quentin Bouniot, Serge Belongie, Zeynep Akata,"The Latent Color Subspace: Emergent Order in High-Dimensional Chaos",arXiv:2603.12261, ICML 2026<pre><code>本工作流使用了 LCS（Latent Color Subspace）节点进行采样期颜色与锐度控制。
ComfyUI 节点实现：facok/ComfyUI-LCS（MIT License）
官方代码：ExplainableML/LCS
基础研究：Mateusz Pach, Jessica Bader, Quentin Bouniot, Serge Belongie, Zeynep Akata,
"The Latent Color Subspace: Emergent Order in High-Dimensional Chaos", arXiv:2603.12261, ICML 2026.</code></pre>LCS（Latent Color Subspace，潜在颜色子空间） 是一套在扩散采样过程中直接操作潜在空间的 免训练颜色与锐度控制方案。LCS is a train-free color and sharpness control toolkit that operates directly in the diffusion latent space during sampling.核心原理：在潜在 patch 空间中，颜色只占据一个 3 维子空间（PCA 捕获 100% 颜色方差）， 其余维度编码结构与细节，与颜色正交。因此可以在不改动结构的前提下推动颜色，也可以沿锐度子空间调节锐度。The core finding: in latent patch space, color occupies a 3-dimensional subspace (PCA captures 100% of color variance) while the remaining dimensions encode structure, orthogonal to color. Color and sharpness can therefore be steered without retraining.技术定位 / Technical position: 与 VAE 解码后的像素级调色不同，LCS 干预发生在采样过程中，模型会在后续去噪步骤中自适应该偏移，因此颜色与光照、反射、材质保持一致。Unlike post-hoc pixel grading, the intervention happens mid-sampling, so later denoising steps adapt to the shift and the color stays consistent with lighting and materials.Latent color control by LCS — community ComfyUI implementation by <code>facok/ComfyUI-LCS</code> (MIT), based on The Latent Color Subspace by Pach, Bader, Bouniot, Belongie &amp; Akata (<a target="_blank" rel="ugc" href="https://arxiv.org/abs/2603.12261v1">arXiv:2603.12261</a>, ICML 2026). Official code: <code>ExplainableML/LCS</code>.节点包以 MIT License 发布。The node package is released under the MIT License.允许 / Permitted:<ul><li>✅ 商业使用（含收费工作流、SaaS、客户交付）</li><li>✅ 修改、再分发、与其他组件组合</li><li>✅ 私有使用，无开源义务</li></ul>MIT 要求的义务 / MIT obligations:<ol><li>再分发时必须附带原始版权声明与许可声明 The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.</li><li>软件按 "AS IS" 提供，无任何担保；作者不对使用后果负责</li></ol>依赖声明 / Dependencies该组件运行时依赖以下库（均为宽松许可，通常 ComfyUI 已自带）：依赖用途<code>einops</code>张量重排<code>safetensors</code>校准缓存读写二者许可均为 MIT / Apache-2.0，不额外增加商用限制。免责与无背书 / Disclaimer and non-endorsement<ul><li>本项目不是 LCS 的官方发行版，与 <code>facok</code>、论文作者、<code>ExplainableML</code> 均无隶属、 赞助或背书关系</li><li>工作流中的参数选择、提示词、输出结果及其合规性，由发布者自行负责</li><li>LCS 的校准缓存按 VAE 权重指纹生成，更换 VAE 后需重新校准</li></ul>This workflow is not an official LCS release and claims no affiliation, sponsorship, or endorsement by <code>facok</code>, the paper authors, or <code>ExplainableML</code>. Parameter choices, prompts, outputs and their compliance remain the publisher's responsibility.本文件用于商用工作流发布时的第三方组件归属与授权声明。 单独发布时请完整保留本文件；打包进工作流压缩包时请置于 <code>CREDITS</code> 或 <code>licenses/</code> 目录内。For commercial redistribution of a workflow that uses these nodes. Keep this file intact, or include it under <code>CREDITS</code> / <code>licenses/</code> when shipping a workflow archive.

mossH3KREA2QW21LTX25_moss6StepsSharpenLCS.json

MOSS 6-Steps Sharpen(LCS)

REDQW21(UNLOCKED) 💴

QW21(DLSS5)LUT NV20-50

QW21加速器(Native acceler)

H3-LTX25-神速重绘⚡

LTX25 Pruned 8+3 Fast

MiniMax H3 Lightx2v 8step

LTX2.3REDMixKrea2

Boogu & Krea2(赤佬版)

Boogu + ideogram 4

Ideogram4 2026.06

DarkBeastZit8🟥SDA WF

🟥Z&Klein9b🟦Hybrid BFS

DBKlein9b🟦 V2.0 BFS

Kolors 2024.1

Files with SHA256: 76bdf954b9a9bbdff0499ef6670addeb0362943c00e7ee5c366720062586e21c