FWIW I'm using an old gtx 1080 Ti to play around, it takes about 21 seconds per image. You can make it go even faster by lowering the timesteps taken from the d

FWIW I'm using an old gtx 1080 Ti to play around, it takes about 21 seconds per image. You can make it go even faster by lowering the timesteps taken from the default 50 (--ddim_steps), though both lowering and raising the value can result in quite different first-iteration images (though they tend to be similar) and seems to guarantee totally different further iteration images (as counted by --n_iter)... I'm with you on the feeling that it's hard to control, whether in refinement or in other ways, but I suspect that'll get a lot better in the next couple years (if not weeks or dare I say days).