As the number of iterations increases, the image quality decreases

[Zhou-Yijie]

I trained the Dreamshaper_v7 model following the settings of train_lcm_distill_sd_wds.py. In the early stages of training (such as 100 iterations), the output image looks like this:

But as the number of iterations increases (such as 1000iteration), the output image becomes low quality and rough, like this:

More iterations did not fix it.
Has anyone encountered such a problem?

[yuyongcan]

Have you solved this?
When I tried to train the SD2.1 model on my own datasets, I found that the loss didn't converge and is higher than the initial iteration.

[Tramac]

+1

[mk322]

+1 I faced the same issue on video generation.

[dragon-cao]

+1 I faced the same issue.Have you solved this?

[yangzhenyu6]

Hello, I also trained the Dreamshaper-v7 model according to the settings of train_1cm_istill_std_wds.comy, but in the end, I got pytorch-lora_ weights. safetensors. How can I call them to see the effect? I tried various methods but couldn't call them. How did you see the effect in the picture?