No original content or contribution.

This code is heavily inspired by FaceFusion and it is completely powered by InsightFace. Do not use this tool, use FaceFusion instead, it is actively maintained, has a cool GUI, and has a good community.

The camera code is taken from Nakul Lakhotia who in turn attributes Miguel Grinberg.

Crappy in this context refers to my code alone, and is no reflection upon the projects I use, it is simply setting the correct expectations.

I saw that FaceFusion can work with video, and it does that by converting all frames to images, and then applies magic to the images.

I wanted to try to:

• Use ffmpeg to get raw frames from the video stream without having to dump to disk in-between.
• Use less and faster magic to try and cut and fill the frame based on what InsightFace crops.
• Use less compute to go towards higher frame rates.
• Be able to run on a webcam feed, though this does not seem feasable with this approach.

No. This barely(and rarely) runs on my machine, and I don't think it is worth the effort to get it to run.

When I run code I do not really care, but friends do not let friends run models with arbitrary code execution. For that reason I have tried what I can to run directly with onnx-model files. I no longer remember where I found the model files I use. The large buffalo network comes directly from InsightFace. They expose this as insightface.utils.storage.BASE_REPO_URL, but I do not like automatic downloads. You can build GFPGANv1.4.onnx from the official pth source using xuanandsix's tool, if you do this care that the input/output names will be different from what crappy-face-swap expects.

SHA256
6548e54cbcf248af385248f0c1193b359c37a0f98b836282b09cf48af4fd2b73  models/GFPGANv1.4.onnx
e4a3f08c753cb72d04e10aa0f7dbe3deebbf39567d4ead6dce08e98aa49e16af  models/inswapper_128.onnx
df5c06b8a0c12e422b2ed8947b8869faa4105387f199c477af038aa01f9a45cc  models/buffalo_l/1k3d68.onnx
f001b856447c413801ef5c42091ed0cd516fcd21f2d6b79635b1e733a7109dbf  models/buffalo_l/2d106det.onnx
5838f7fe053675b1c7a08b633df49e7af5495cee0493c7dcf6697200b85b5b91  models/buffalo_l/det_10g.onnx
4fde69b1c810857b88c64a335084f1c3fe8f01246c9a191b48c7bb756d6652fb  models/buffalo_l/genderage.onnx
4c06341c33c2ca1f86781dab0e829f88ad5b64be9fba56e56bc9ebdefc619e43  models/buffalo_l/w600k_r50.onnx

My code falls under MIT License but the models you'll need from InsightFace are as they say "non-commercial research purposes only" which means this code is worthless on its own, except as a tool for understanding.