Created this workflow to improve the performance of the Reactor Fast Face Swap node since it tended to glitch out quite a lot when handling more complex faces.
The vast amount of time is spent waiting for reactor to do its thing with the face boost on but overall i'm happy with the result.
For the vae loader you can use whatever vae you wish, the purpose for it is to re-encode all the images coming out of the Masking Helper to float32 from double.
Models to download:
* GPEN-BFR-512.onnx -> models/facerestore_models/
There's a reddit thread about it here: https://www.reddit.com/r/comfyui/comments/1kno9i4/the_ultimate_productiongrade_video_photo_face_swap/