What causes black spots on the face of images generated by models trained on my customer dataset？

[WYanNuo]

What causes black spots on the face of images generated by models trained on my customer dataset？

[WYanNuo]

I used wav2lip_train.py for training. What I meant was that black dots appeared on the forehead of the face generated by the model, and black dots also appeared on the face during inference. I don't know how to avoid this problem.

[zhb1991nm]

Hello, I've encountered the same issue. Have you found a solution?

[jinwonkim93]

me too but mine is green.

[killnice66]

mine was blue

[jinwonkim93]

@killnice66 Were you able to converge syncnet to 0.25 under?

[killnice66]

@killnice66 Were you able to converge syncnet to 0.25 under?

nop, mine syncnet is about 0.30 , my model is wav2lip288

[hughsando]

It might be a overflow issue. You could try clipping before converting to uint8, np.clip( value*255 , 0, 255 ).astype(np.uint8) for the output values.

[sd707589]

Keep training until your L1 loss is below 0.008. Then the artifacts may disappear.