So, this may be more of a conceptual PR than one that can be directly merged, but here's a list of the changes and modifications I've made:

1. Support loading weights directly from a v1.5 .ckpt or .safetensors file, falling back to the "stock" weights otherwise.
2. Add a prompt field. Does it make a difference? IDK. But it's there for funsies.
3. Add an option to use FP16 or not.
4. Fix multiple pathing issues when run on Windows due to slashes.
5. Add an xformerify method to enable xformers on various modules, falling back to attn2.0 if not found.
6. Update requirements.txt to install CUDA vartiants of Torch, fix other dependencies (Triton is not available on Windows)
7. Various other generic code cleanup.