Open aagainbeyond2 opened 3 months ago
Hi, thanks for your interest. We can't promise immediate release of the training code. However, it's based on the diffusers library for fine-tuning SD v1.5 and the hyperparameters used are described in the paper, so it shouldn't be hard to reproduce it.
@foivospar @aagainbeyond2 i am trying to reproduce it and it seems to be working-ish but also so hard to tell without having the same exact dataset (it also takes an insanely long amount of time to converge, 15k steps at bs=64 with lr=2e-4 it is still so far off but compared to probably 5 epochs on 42M + 15 epochs on the HQ it is not tht much).
Have you been able to reproduce the training successfully? If so, could you please share it?
+1
would you release the training code? thank you