mlpc-ucsd/TokenCompose

fine-tune

Closed this issue · 4 comments

Mian2001 commented

For fine-tuning Stable Diffusion v1.4, around which step does it typically converge? This information would be helpful for my further research work.

JamesSand commented

Hi
As indicate here, according to our experiments, SD 1.4 converges around 24500 steps.

Mian2001 commented

Thank you for your response. I mean the fine-tuned SD v1.4, without using TokenCompose. Does it also converge at 24,500 steps?

JamesSand commented

We are not sure about for the convergence step for just finetuning SD 1.4.
From our side, we first find SD 1.4 finetune with Tokencompose will converge at 24500 steps. So, in our paper, we also choose 24500 to finetune SD 1.4 without Tokencompose.

Mian2001 commented

OK, thank you.