Finetuning & result about ddad HOT 1 CLOSED

Hello, appreciate your inquiry. Throughout the training process, we encountered significant instability with diffusion models. Consequently, we have shared the checkpoints, accessible in the readme file. So you can easily download them and use them.

Typically, models employing diffusion models benefit from a large batch size for effective model guidance. Adjusting the batch size to 16 is expected to yield different results. However, due to resource constraints, we have set it to 32.

In cases of limited resources, alternative techniques like gradient accumulation can also be employed.

I hope this explanation adequately addresses your question.

from ddad.