search

deepfakes / faceswap-model

Tweaking the generative model
147 stars 133 forks source link

Paper & source: progressive growing of gans #25

Open b4nn3d opened 6 years ago

b4nn3d commented 6 years ago

They officially recently ported this in tensorflow, results are 1024*1024 photorealistics images.

paper (nvidia research) paper (arxiv) source video

btw, they train with telsa V or P (16gb vram), I guess we should tuning down to 256 or 512. But the idea of progression of gans its really cool (and time saving in training)

iperov commented 6 years ago

minimum req: 16GB of video ram.

b4nn3d commented 6 years ago

Only if you go at 1024*1024 resolution.. in the official repo there is a benchmark tool, you can try by yourself

iperov commented 6 years ago

but how it can be used in face swap task ?

bryanlyon commented 6 years ago

This is easy to use in face swap. While yes, as it stands this paper's results would need to be tweaked to fit. Right now, faceswap's GAN is broken anyway. I'll actually be at CVPR and will be attending this talk so I'll see if I can ask them some questions on this topic.