Sana
Sana copied to clipboard
Question on SCM timestep * 1000 for SANA-Sprint
In the paper, it is discussed that SANA-Sprint disable 1000t in training, which leads to better training stability.
However, in the code, SANA-Sprint still seems to keep the 1000t? I am just wondering because I am attempting to fine-tune the diffusers checkpoint, and I am wondering if I should keep the *1000.
https://github.com/NVlabs/Sana/blob/daa661fd85ffa881983d760006f3ea8598058c1f/diffusion/model/nets/sana_multi_scale.py#L426
Thank you very much!