Significantly improved img2img quality by scaling masked area before inference

[TingTingin]

With this experimental addition you can increase the quality of img2img generations by pre-scaling the masked area

E.G
prompt: elon musk face
This is the selection mask

This is the image generated from using regenerate only masked area

This is the image generated with the new resize and regenerate only masked area

As you can see the quality is much improved

Potential Improvements

• Upscaling the image using realesrgan or another upscaler yields even better results as of now the image just uses LANCZOS scaling
• The mask cropper currently "zooms in" on the mask then extends a *square from the left-most side of the image that encapsulates the entire mask it would be better if it was centered would also potentially help with *blending
• Because of the crop on the image certain context is lost of the overall image it would be nice if there was a Context setting that dictated how much to expand the sampled area (at the cost of quality) to increase coherency
• the code currently uses a global variable returned_info to pass some data around would be better if it was handled more efficiently
• *square I was under the impression that you could send any 64 multiple dimensioned image to get it generated so the initial code wouldn't use squares but instead would use whatever shape would fit while having a 64 multiple dimension on both sides up to the max width/height that was specified in the ui but I kept getting an error so I had to change the code to always create a square instead not sure what I did wrong
• *blending The blending needs to improve as of right now it takes the generated image applies the mask with a blur and call it a blur but there's a few issues
• • when you blur an edge you need to have enough "room" or else the blur will go over the edge and give a hard line essentially the mask blur should also increase the size or "room" the mask has to accommodate the blur more appropriately and of course also make the mask is the same size as the image its getting applied to at the end