get blurry image after lora finetune

[hanwenxu1]

I'm trying to use train_cogvideox_lora.py to finetune 2b model with customized dataset. When I set the train lr is 1e-5, and the epoch is 2, I get some normal result. Then I found some advice that the lr of lora should be larger, like 5e-4 to 1e-3. I tried to set lr as 5e-4 and the epoch is 10. But I get worse results. I wonder that why it became worse and it seemed so vague. The train loss is here: