search

THUDM / CogVideo

text and image to video generation: CogVideoX (2024) and CogVideo (ICLR 2023)
Apache License 2.0
9.29k stars 874 forks source link

sampling is different between diffusers and sat ? #482

Open lith0613 opened 1 week ago

lith0613 commented 1 week ago

Hello, I am currently encountering an issue where the effect becomes noticeably blurry after using conversion code to transform a SAT training model into diffusers. I have pinpointed that the model parameters are aligned, but there is a clear discrepancy in the output of the scheduling part. Specifically, the output after executing scheduler.step in the diffusers code does not align at all with the code in the SAT version’s sampling.py file.

zRzRzRzRzRzRzR commented 1 week ago

Regarding the 2B version, I think there might be some differences. Perhaps @a-r-r-o-w could help me with an explanation of this part?