light-and-ray / sd-webui-resynthesizer-masked-content

Adds Resynthesizer - not NN old open-source content-aware-fill algorithm as masked content in stable-diffusion-webui

Resynthesizer as masked content

This extenstion for AUTOMATIC1111/stable-diffusion-webui adds new value of "Masked content" field in img2img -> inpaint tab. It uses Resynthesizer - a very old (2000 year) open-source equivalent to Adobe Photoshop's "Content-Aware Fill" feature

If you install lama cleaner, "Resynthesizer" will appear as model in extras tab

Generally it's worse and slower then lama, but in a few scenarios if lama has some concepts which it can't remove, resynthesizer can help. Also it can on hard for surfaces e.g. grid, sand

Example of concept which lama cannot remove:

mask

lama

resynthesizer

Links:

• python lib: https://github.com/light-and-ray/resynthesizer-python-lib
• c implementation: https://github.com/61315/resynthesizer
• original project: https://github.com/bootchk/resynthesizer

If you have problems with windows build, I would be thankful if you can fix it and make PR in resynthesizer-python-lib, because:

For .dll I've used x86_64-w64-mingw32-gcc from sudo apt-get install gcc-mingw-w64 instead of gcc. It shows strange worse result on my virtual machine, I don't know why. They are the same if I build in VM using clang, or if I build it on Linux