Javier Solís García

@Dart120, which solution was the final one? I have already tried using the AdamW and the learning rate suggested by @volcanik. Still, while I have seen an improvement in the...

Thank you very much for your answer, @Dart120; I'll try to apply this information to check if I can improve my results. Good luck with your coursework!

I have finally implemented what you said @Dart120, and it works as expected. Thank you again! For anyone reaching this thread in the future, the scheduler of [diffusers](https://github.com/huggingface/diffusers.git) library seems...

Maybe, to be more consistent with the implementation, c_in should be updated to be: `1/((sigma - sigma_min)**2+sigma_data**2)**0.5` Any thoughts about that?

@Kinyugo @wubowen416 In addition, I have recently discovered something interesting. Although I am almost sure that the Consistency models and Improved Techniques for Consistency Training did not mention anything related...

Another hint, even the recent paper of consistency models made easy use this two rescaling factors for the noise and network input: https://github.com/locuslab/ect/blob/4311059770f54821d151a9b0e1f76770a5f3930e/training/networks.py#L700-L718