Comment by Jack000

Depends on the dataset. You can probably get decent results by restricting the modality of the images (faces, cars, bedrooms etc)

I trained from scratch with 4x3090 and while it’s not as good as SD it’s surprisingly better with hands.