Request for image enlargement about stable-diffusion-webui

[raldnor]

Perhaps this can be incorporated?
https://github.com/jquesnelle/txt2imghd

[AUTOMATIC1111]

raldnor: rejoice, for your prayers has been answered

[1blackbar]

This is quite incredible !!! ID love to be able to upscale like this in GUI, this result is amazing , so it splits the image into parts and upscales every piece then merges,, its pretty smart, i think midjourney has similar stuff with kind of a grain texture to get rid of artificial digital look
https://raw.githubusercontent.com/jquesnelle/txt2imghd/master/gallery/00091ud.png
Its quite different cause its using img2img to upscale further but it splits the image cause jamming full big one into img2img to will wipe ram space immediately.
So first it generates txt2img result , then realsrgan upscales it and then img2img adds details to each part of the 512 chunks of full image... i wonder if you could aadd extra step to upscale face with gfpgan and then run img2img on top of that so its txt2img - realsrgan - gfpgan- img2img .
But maybe gfpgan upscale before img2img would look worse than adding it separately at the end, im not sure.
But i think it would have to be able to be turned off if you dont want it to upscale everything.

Or... maybe trying to lookup his code how hes using img2img to upscale in chunks of 512 images would be better idea ? So there would be separate img2img upscale button ? This way you could add gfpgan sooner or later.More modular.
I think hes upscaling in img2img using the same prompt and seed so it needs to be done right away. unless it would be possible to reacquire prompt and seed from the exported image and upscale would work fine.

[AUTOMATIC1111]

There's already Cupscale, I can't imagine how a tab would be more convenient than that.

[raldnor]

As far as I can tell Cupscale is Windows only, I run SD with the webui in Linux and intend to package it into a docker container. I thought it would be nice to have all functions in one package.

[AUTOMATIC1111]

1blackbar: I have the reddit's script, have not played with it yet, it's a different feature request

[raldnor]

automatic1111: You're the best! Thanks a lot!