If input images didn't need resizing because they were already smaller than max width/height then they didn't get normalized to a multiple of 64. This caused an exception like the following:
```Sizes of tensors must match except in dimension 1. Expected size 4 but got size 3 for tensor number 1 in the list.
```
while the previous version did produce much better blending it also makes images that lack detail for some reason.
tests: Added more tests to help catch this sort of thing earlies
fix: found that median blur is really slow, so I made sure we only do it on downsampled masks. Was taking like 3 minutes to run on the large pearl girl picture on M1
- docs: update examples
- 🎉 fix: inpainted areas correlate with surrounding image, even at 100% generation strength. Previously if the generation strength was high enough the generated image
would be uncorrelated to the rest of the surrounding image. It created terrible looking images.
- fix: mask boundaries are more accurate
Specify advanced text based masks using boolean logic and strength modifiers. Mask descriptions must be lowercase. Keywords uppercase.
Valid symbols: `AND`, `OR`, `NOT`, `()`, and mask strength modifier `{*1.5}` where `+` can be any of `+ - * /`. Single-character boolean
operators also work. When writing strength modifies know that pixel values are between 0 and 1.
- feature: apply mask edits to original files
- feature: auto-rotate images if exif data specifies to do so
- fix: accept mask images in command line
If the x_sample was a bfloat on the gpu but the cpu doesn't support bfloat, that can cause a TypeError
```
File "/home/stdiff/.local/lib/python3.10/site-packages/imaginairy/api.py", line 292, in imagine
x_sample.cpu().numpy(), "c h w -> h w c"
TypeError: Got unsupported ScalarType BFloat16`
```
Seems to be caused by incompatible types in group_norm when we use autocast.
Patch group_norm to cast the weights to the same type as the inputs
From what I can understand all the other repos just switch to full precision instead
of addressing this. I think this would make things slower but I'm not sure. So maybe
the patching solution I'm doing is better?
https://github.com/pytorch/pytorch/pull/81852
