SD 2.1 training is not working as expected

AI-Hypercomputer / maxdiffusion

Apache License 2.0

149 stars 17 forks source link

SD 2.1 training is not working as expected #73

Closed terrychahn closed 4 months ago

terrychahn commented 4 months ago

When using the base21.yml script to train a SD 2.1 model, the generated checkpoint does not appear to be trained from the dataset correctly. The output images are the same as the base model. It has also the same output after training with modified params of 15000 training steps and a learning rate of 1e-5.

entrpn commented 4 months ago

Fixed through #70 - mainly keeping weights in fp32. Closing this issue, but feel free to re-open.