pyphan1
pyphan1
This is happening because you are using a non-inpainting model to do inpainting, inpainting should use the custom inpainting model and not the normal text-to-image or image-to-image model.
@luser350 this is often caused by a high learning rate or a very low batchsize, it is an area of confusion where gradients must have exploded below 0 or higher...
if you still have this problem, the fix is to either install the extension, which didnt work for me on windows, or use this code workaround with the same functionality:...
show me an example
it seems that gfpgan didn't run at all in your case, did you verify that it ran without errors?
probably because this is the most critical layer, it converts the output of the convolutions to a class probability so it needs high precision
maybe because the liveness detection model crops and aligns the face first? if so then the only factors that can be used to determine if it's a flat picture is...
@Yevrey921 did you ever fix this problem? does gtx1650 work without --no-half and how many it/s do you get?
@janeyx99 yes that's exactly it! it would solve the unnecessary high vram usage when training GANs
update: seems that adding xm.mark_step() before and after inference makes it work somehow, and it's fast