face swap instead of pet swap
ak9250 opened this issue · comments
https://github.com/shaoanlu/fewshot-face-translation-GAN
uses modules from FUNIT and SPADE for face-swapping but the quality of swaps isnt as good as FUNIT, are there particular limits like expressions and rotations?
I just ported funit to pure keras and started training with 2 classes for face swap
part of dfl
when approx it will appear in your dfl?
would love to try it
[dfl looks a bit overloaded by end-user convenience features, cleaner task-oriented code would b perfect]
In the last page of the FUNIT arxiv paper (which will be presented in ICCV 2019), we do have the few-shot face translation experiments. We simply use CelebA for the experiments. I expect combining with SPADE and utilizing landmarks should lead to better performance thought.
@mingyuliutw I am trying to train FUNIT on celebA but I am confused. What are the classes? If I take every person as a different class then there would be too much classes. Or can I take real as 1 and fake as 0, nothing more? Please Guide me.
Each human identity is a class. CelebA has the person name for each photo. You can divide the training set into different classes using the name.
I trained FUNIT on 256 VGGFace dataset clases for a week.
Currently, it seems that it cannot correctly convert unknown persons.
@mingyuliutw, are 256 persons with 69k total photos enough for the model?
should I expand to 1024, or 2048 persons ?
@mingyuliutw can you give me advice for network config for 1024 person classes and 200k photos?
@mingyuliutw is it good that one class has 60 photos and other class has 500 photos?
@iperov
would you like to share me your keras code? thank you.
thank you ~~
@iperov
Hi I'm quite interested how you trained fuint for face swap
Note N faces in the trainig data set, you just followed the origin fuint traning set and treat N ids as N classes right?
for 1024 persons I got unsatisfactory result.
seen faces swap:
unseen faces swaps are unrecognizable:
May be if train it on 8000 persons on bigger size funit we will get better result, but I have no time and no hardware for that.
for 2 persons result is much worse than classic deepfake autoencoder.
