Skip to content

Guideline to reproduce experiments on FLUX.1-schnell #3

@Leong1230

Description

@Leong1230

Hi again!

I am currently exploring RL on distilled models, and your work has been highly inspiring. I am looking to reproduce your experiments on FLUX.1-schnell (a 4-step distilled model), which I noticed wasn't explicitly covered in the Flow-GRPO repository. However, I couldn't find the corresponding hyperparameter configurations or launch files for this specific setup in this repo. Could you kindly provide some guidelines or share the config files for training on such few-step models?

Additionally, have you noticed a gap in the RL training dynamics between 50-step models and few-step models? My recent experiments on our own 8-step model show that RL on few-step models suffers from a severe and rapid image quality degradation (approaching catastrophic collapse) after only 100~200 gradient updates. This occurs even after I intentionally omitted the $\sigma_t$ term in the loss calculation.

Thank you!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions