feat(ui): upscaling tab

[maryhipp]

Summary

A new tab for multidiffusion upscaling ✨

• Requires installed spandrel image-to-image model and tile controlnet for the base model used in multidiffusion step

Related Issues / Discussions

QA Instructions

Merge Plan

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated (if applicable)
• Documentation added / updated (if applicable)

[psychedelicious]

I fixed a couple things and broke others.

I "fixed" a problem where scale would be null if the upscale model used didn't have a specific pattern of x then a number. I "fixed" it by removing that parameter. This breaks the size preview.

Also removed the resize node, because it wanted scale. This breaks when the image size doesn't end up as a multiple of 8.

I'll propose an alternative way to handle the initial upscaling logic - create a new node:

• Accepts image, model, tile_size, scale: float, overshoot: float, fit_to_multiple_of_8: bool
• It runs the model until upscaled_image.width >= scale * overshoot, then uses a simple resize to the nearest multiple of 8 of the scale.
• overshoot = 1 means you do minimal upscaling
• Higher values let the upscaler generate more detail, which results in finer details after scaling back down, which I found yields better results.

This approach also reduces the number of intermediate images created, which improves perf - a 8k image is about 100MB and will take many seconds to save.

It also means you can precisely tune how large you want your output to be, and eliminates the need to figure out upscale model's scale factor.

[maryhipp]

@psychedelicious Is it not safe to assume that these models will have x# in the name? Or just default to 2 if it doesn't?