Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

args.noise_offset is not working for flux training? #1861

Open
ziyaxuanyi opened this issue Dec 31, 2024 · 2 comments
Open

args.noise_offset is not working for flux training? #1861

ziyaxuanyi opened this issue Dec 31, 2024 · 2 comments

Comments

@ziyaxuanyi
Copy link

When I check the latest version of the code, I found that args.noise_offset is not working for flux training.
Why?Will it cause bad results?

@67372a
Copy link

67372a commented Dec 31, 2024

@ziyaxuanyi offset noise and multires/pyramid noise are both "bandaid" attempts to address the issues (inability to generate very dark or very bright images) with the noise scheduling under typical noise scheduling used by SD and SDXL. It isn't required and doesn't have benefit to v-pred + ztsnr, or flow matching (which is what Flux and SD3 use). See https://arxiv.org/abs/2305.08891 for details on the issues with typical / epsilon noise scheduling.

@ziyaxuanyi
Copy link
Author

偏移噪声和多重分辨率/金字塔噪声都是“创可贴”尝试,用于解决 SD 和 SDXL 使用的典型噪声调度下的噪声调度问题(无法生成非常暗或非常亮的图像)。它不是必需的,对 v-pred + ztsnr 或流匹配(Flux 和 SD3 使用的)没有好处。请参阅 https://arxiv.org/abs/2305.08891 有关 typical / epsilon 噪声调度问题的详细信息。

Yes, you are right. I have modified the code to make it effective when training flux, and tested the generation effect of the training model. It has indeed deteriorated.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants