Train fine, but merge / convert not about faceswap HOT 5 CLOSED

I don't know much about this issue, but I think it comes from a bad training set of images. Please note that if you have too few images or too few diversity in images, the model cannot invent angles he does not know about. If you think it is not the case, feel free to reopen

from faceswap.

Did you find a solution to this problem with FakeApp?

I'm having the same issue, previews during training look great with loss less than a 0.01, but the merged/converted pics look like a monster.

[Links Removed]

from faceswap.

Well your input data is pretty crappy. That’s probably what’s wrong. There’s a lot of noise and it’s blurry.

from faceswap.

I'm using 250 frames extracted from a very short video for A, with subject A looking only straight forward and down, to keep things simple for a first test.

I'm using 1500 frames extracted from a longer video for B (5000 original frames, of which I removed 3500 that were not matching orientation or expressions of the 250 in video B).

Each and every single of the 250 aligned crops for input A is converting well in the previews, but not in the merged result.

[Links Removed]

NOTE: I think it's an issue with FakeApp 1.1. Since my first post above, I tried using the open-sourced OpenFaceSwap, and results are almost as good as the previews, very satisfied result for a first (now second) test :)))

from faceswap.

Have you tried it with the current version of this repo?

from faceswap.