fsgan

[ak9250]

thoughts on fsgan and there method for face swapping, results look impressive https://nirkin.com/fsgan/

[iperov]

Nothing impressive.
fsgan's faceswapping is much worse than existing methods.

Reenactment at 1:23 in the video https://youtu.be/BsITEVX6hkE?t=83
produce unstable background.
My new avatar model in DeepFaceLab produce more stable background https://www.youtube.com/watch?v=rRGOQpR_3hw

[ak9250]

@iperov good results, is your model few shot or does it have to be trained for each face pair?

[iperov]

@ak9250 not few shot. As regular deepfake.

[gstark0]

Waiting to see the code release, looks promising except that unstable background which is awful.

[lseiyjg]

Unstable background can be ignored, because it usually uses a face mask to swap the face. FSGAN is quite good as its demo shows.

[iperov]

@lseiyjg no, in reenactment approach background must be stable, because reenactment is not a face swap.

[szad670401]

@iperov The background can be removed by face landmark.
No Few Shot methods like DeepFake may only help little in some instant environment.
Some application scenarios that have no time to wait for training.

[iperov]

@szad670401 bla bla bla. Have you anything to show?

[szad670401]

Are you stupid?
Have you anything to show?
Is that the one modified from deepfake and used it to fool 10k stars?