[Feature Request] Improve flexibility of SD Upscaler

[ThereforeGames]

I think the idea of an SD upscaler is one of the less explored areas of this tech - there's so much untapped potential here.

Imagine if we could draw boxes inside parts of the image and give each box a slightly different prompt or denoise strength. This alone could massively improve the quality of the results.

Another user suggested a batch size option: #169

Taking this a step further, I would love the ability to click on a "chunk" inside of the upscaled image to regenerate only that piece.

Or how about exploring different methods of blending the tiles together? I'm not sure exactly how the current implementation works, but the tile fading can be a bit obvious sometimes. Are these overlaid parts processed through img2img at the end? Maybe a second pass could help.

Alternatively, there are ways of possibly feeding the original image back into the result - in Photoshop, you can do a high pass and a little surface blur on the original image, set it to Overlay mode and lower the opacity, place on top of the upscaled version... this brings back some of the original's identity while preserving the clarity of the upscale. Would be interested to see this sort of technique directly inside of an SD workflow.

There are a lot of possibilities here. Thoughts?

[gibsonfan2332]

I second this! This feature has so much potential if it can start getting better results and less faded objects. It would also be nice to get some examples of people successfully using it and what settings they used. I am trying out lots of settings myself, but it is a very slow going process.

I have yet to find anyone show good results and all the settings they had.