huggingface/diffusers

[Flux ControlNet] Add support for de-distilled models with CFG

Open

#9,635 创建于 2024年10月10日

在 GitHub 查看
 (11 评论) (0 反应) (0 负责人)Python (22,190 star) (4,562 fork)batch import
contributions-welcomegood first issue

描述

Flux with Inpaiting and ControlNet currently yields bad result with the base model.

To echo this comment, using de-distilled models could potentially help getting better outputs.

Currently the Flux ControlNet pipelines do not support models like https://huggingface.co/nyanko7/flux-dev-de-distill

贡献者指南

技术栈
pythonpytorch
领域
machine learning
议题类型
feature
难度面向新贡献者的预计实现难度，1 表示很小改动，5 表示专家级工作。
3
预计时间有经验贡献者完成调查、实现、测试并准备 pull request 的粗略时间范围。
1-2 days
活动状态议题当前的可参与程度：新鲜、活跃、陈旧、阻塞或等待维护者输入。
active
清晰度议题是否清楚说明期望改动、验收标准和下一步。
mostly clear
前置要求
Understanding of Flux ControlNet pipelineKnowledge of CFG and model distillationFamiliarity with diffusers library
新手友好度1-100 的估计分数，表示该议题对首次贡献者的友好程度。
30
研究方向
Investigate the current Flux ControlNet pipelines in the diffusers repository to understand how models are loaded and used. Refer to the linked comment in PR #9571 and the de distilled model by nyanko7. Identify where CFG is applied and why de distilled models are not supported. Then modify the pipeline to support loading de distilled models, ensuring compatibility with CFG by adjusting the noise prediction logic accordingly.