モデルを手直した感触だが、画像が多いとprodigyの結果は良くない。枚数が多いなら、Lionが結果が良い。
今回の例:
--optimizer_type="Lion8bit" --lr_warmup_steps=100 --max_grad_norm=1 --text_encoder_lr=0.00015 --unet_lr=0.00015
I have a feeling that I reworked the model, but if there are a lot of images, prodigy does not give good results. If there are many images, Lion has better results.
Example of this time:
--optimizer_type="Lion8bit" --lr_warmup_steps=100 --max_grad_norm=1 --text_encoder_lr=0.00015 --unet_lr=0.00015