Hope

> Not related to AMD The `use_8bit_adam` problems potentially are, as bitsandbytes includes a C extension which wraps some CUDA functions, i.e, doesn't run through pytorch-rocm. Not really anything that...

Gamechanging or not, it's unlikely to catch on if hobbyists can't play around with it, so the fact that this technique requires a workstation class gpu is a huge downside...

@Desm0nt The saying that comes to mind is "Beware trivial inconveniences". By which I mean adoption of two competing things is heavily influenced by which one is more accessible. Being...

@ShivamShrirao If I'm reading things right, 8bit AdamW should be a drop in replacement and the modified CrossAttention class seems like it should just be able to replace the one...

@rinongal Styles seem generally doable, but characters suffer from the problem where their simplicity means a single misplaced line or shape is enough to make them appear deformed or unrecognizable,...

Looks like they're doing some sort of face swapping/inpainting rather than generating the whole image from scratch

>This would be great to have native in AUTOMATIC1111 to make it more accessible. Dreambooth requires just under 24GB with all optimizations. Ease of use isn't the problem, the raw...

It's now down to 12.5GB (but for diffusers, so not usable with this repo anyway). Yet I'd still argue that *training* a model is an entirely separate use case from...

Models in the diffusers format aren't currently supported afaik. Either support would have to be added or the dreambooth optimizations would have to be ported to the CompVis-based implementation