Open wen313 opened 1 month ago
It's interesting, because I did some tests weeks ago with the flux-dev trainer and I noticed that the program was downloading files from both FLUX.1-dev and FLUX-1.schnell. But now, I updated the repo, did some tests and now the program is downloading only FLUX.1-dev files.
I checked the commits and the issue seems to be fixed in the commit 4aa19b5c1ddf30aad233e907fdd45447878a2231. But, from my last tests, I'm getting a different quality of my LORAs and I needed to revert back to the previous commit.
If you compare the files between FLUX.1-dev a FLUX.1-schnell excluding the transformer folder, the only folder that changes is the scheduler. So, previously, the trainer was using the scheduler from FLUX.1-schnell, and now its using the scheduler from FLUX.1-dev.
Will the different configuration in the scheduler affect the training? If thats the case, does that mean that the training was wrong previously and now it should be as expected? Because it seems that people were having good results with the FLUX.1-schnell scheduler on FLUX.1-dev. So, I wonder if the trainer should keep using the scheduler from FLUX.1-schnell.
Just curious.
https://github.com/ostris/ai-toolkit/blob/3400882a8099645ce4c797f57ac258f1e1424ffd/toolkit/stable_diffusion_model.py#L535
The processing here is different from is_v3 above, which causes the text_encoder folders to download FLUX.1-schnell instead of the actual configuration. Can we also add:
I also want to confirm whether the text_encoder files in FLUX.1-schnell and FLUX.1-dev are the same. Will using these files in FLUX.1-schnell affect the training results?