[Feature] support for realistic upscale model

[randomqhacker]

Feature Summary

Support a realistic upscale model like realesr-general-x4v3.pth

Detailed Description

Support for a realistic upscale model would really help make diffusion on CPU/iGPU more useful. There is already support for an anime upscale, but it is not useful for most other styles of artwork or photos.

https://github.com/xinntao/Real-ESRGAN/releases/

Alternatives you considered

Images can be scaled independently of sd.cpp, but this prevents direct use of sd-server and requires additional libraries and programs.

Additional context

My use case is using the sd-server API from another web application.

[wbruna]

The documentation is a bit outdated: #885 added support for more models. Not sure if realesr-general-x4v3 is one of them, though.

I keep a few that I've tested myself at: https://huggingface.co/wbruna/upscalers-sdcpp-gguf .

[randomqhacker]

Thanks @wbruna, I can confirm those models you linked work in sd-cli. I tried to use your instructions for converting realesr-general-x4v3.pth but it is in a different/newer format that is not supported.

@leejet you can consider this issue almost closed but maybe update https://github.com/leejet/stable-diffusion.cpp/blob/master/docs/esrgan.md to indicate that these other models/finetunes are also supported? Thanks for all your hard work!

[leejet]

Thank you for the reminder — I’ll update the documentation.

[rene-descartes2021]

I was in similar boat. It was hard for me to find a working 2x one, most realistic ones I tried did not work. Here are the two 2x ones I found that did work following wbruna's script:
https://openmodeldb.info/models/2x-ESRGAN
https://openmodeldb.info/models/2x-PSNR

I guess they're both original ESRGAN models converted to different scale.