search

tencent-ailab / IP-Adapter

The image prompt adapter is designed to enable a pretrained text-to-image diffusion model to generate images with image prompt.
Apache License 2.0
5.08k stars 331 forks source link

loss hard to decrease #253

Open renrenzsbbb opened 8 months ago

renrenzsbbb commented 8 months ago

Thanks for your great work. The result is surprising. So, I try to train or finetune your model using danbooru2000 dataset. Unfortunately, I find that the loss hardly decrease no matter train from scratch or finetune your model. I train it with batch size 16 in 2 * V100 GPUs. And then, I add a controlnet, the loss is also hard to decrease. Can you tell me your loss trend of change and give me some advices? Below is my tensorboard. Thanks in advance. tensorboard

xiaohu2015 commented 8 months ago

I think you should generate some visual images in your training Because loss is not a reliable indicator