Nerogar
commented
1 week ago It's now fixed for flow matching models. And I don't think it was ever broken for normal diffusion models
Closed dxqbYD closed 1 week ago
Nerogar
commented
1 week ago It's now fixed for flow matching models. And I don't think it was ever broken for normal diffusion models
What happened?
Parameters such as "minimum noise" are documented to focus training on image composition instead of details, but actually do the opposite.
Compare the setting in the first screenshot with the actual timestep distribution used during training:
The latter screenshot is a histogram of model_output_data['timestep']
What did you expect would happen?
Mirrored
Relevant log output
No response
Output of
pip freezeNo response