search

shaoanlu / faceswap-GAN

A denoising autoencoder + adversarial losses and attention mechanisms for face swapping.
3.38k stars 844 forks source link

1 & a half day of training... #51

Open ZeroCool22 opened 6 years ago

ZeroCool22 commented 6 years ago

Why the new face looks so out of place? (i'm very glad with the color correction).

https://www.youtube.com/watch?v=sKhBkUVeFyQ

shaoanlu commented 6 years ago

You can increase the coverage of bounding box. In section 13, def process_video(input_img), there are some hard coded values for ROI adjustment: h//15 and w//15. To increase the coverage, we simply flip the positive/negative sign before w//15 and h//15. e.g., x0+h//15 becomes x0-h//15 and so on.

But be careful, don't modify the w//15 and h//15 in def get_mask(roi_image, h, w).

I'm working on a more elegant method to address the out of place problem. However I can not reproduce such issue on my dataset so it's progressing slowly...

ZeroCool22 commented 6 years ago

Check your google mail, i send you (complete data set + Input Video) so you can reproduce the problem.

shaoanlu commented 6 years ago

Awesome! I'll look into it.

petergerten commented 6 years ago

I have the same issue. I thought this is an alignment problem !?