Implement SimSwap (face-swapping with one single trained image) in Stable Diffusion

[ivaxsirc]

When you train faces in Dreambooth, the eyes mainly look very bad in stable diffusion, i'ts true that Codeformer and GFPGAN repair faces, but sometimes the result of the face has nothing to similar with the real face.

If once you generate the image you use impainting, it improves, but the desired result is not achieved either.
If there were some way to implement SimSwap, I think the results would improve ostensibly, since that face that you already have in a photograph could be applied to the image generated by artificial intelligence.
Could you consider incorporating it?
Thank you very much for your dedication and your work.

[Elevory]

Prior to Stable Diffusion, the best results I got for faceswapping was image -> sber-swap -> simswap -> GPEN. This can produce a perfect swap almost every time, but bear in mind that sber-swap and simswap have relatively long inference times.

Would be cool to have some of those options baked into the UI.

[fuselayer]

Wait, so you fed the image to seber, then faceswapped the sber output using simswap, then finally faceswapped the simswap output with GPEN?