Overfitting?
#58
by
tim0
- opened
I have done two training runs, one with the v1.5 model and one with the v2.1 768x768. Both times the models seemed to overfit heavily, making it very difficult to get images in a diferent style than the ones I trained with (photos of my face in different settings / different facial expressions).
Training locally with the dreambooth extension for the AUTOMATIC1111 stable diffusion webui gave far better results
I experience a similar thing than you for v2-1-768, thought v1-5 works quite nicely for me. What hyperparameters did you use for AUTO1111 extension?