high resolution

[ClashLuke]

Quite a few things are happening here. We'll probably have to clean up the code, but the concepts are simple.

1. Instead of using .repeat_interleave() and .expand(), this PR uses only .expand() greatly reducing memory usage (this change alone allows 3072x3072 on one 3090)
2. A simple "size" argument for the generator
3. Warmup / Cooldown for guidance factor instead of a pure constant factor
4. CFG only when using FABRIC (perfect when using it only on the first half of the steps, which reduces runtime cost from (2+K)F to (2+K)F/2 + F while improving image quality noticeably)
5. fused forward pass of positive samples and batch, removing the entire "hidden state caching" and CPU offload business, shaving off another 5% - 10% of runtime while halving memory usage in my one test

Finally, this PR removes negative images. As they are an intended feature, adding them back into the code might be required. (Not difficult, just tedious.)

[ClashLuke]

(3) and (4) together allow the creation of 1024x1024 images as below

compared to the unguided baseline

and default fabric

[dvruette]

Very cool, thank you! Already looking great, love the optimizations and high-res images.

As far as what's required to merge: Yes, indeed, negative feedback is kind of an integral feature so we should try and add back support for that. Also, if possible, we should avoid requiring the xformers dependency, especially when nowadays PyTorch 2.0 is pretty much the norm and most people have access to F.scaled_dot_product_attention.