search

Heonozis / FaceShifter-pytorch

115 stars 27 forks source link

Have you ever obtained similar performance as in the paper? #1

Closed JeyesHan closed 4 years ago

JeyesHan commented 4 years ago

Thank you so much for your brilliant work. I am wondering the performance of your implementation currently. Have you gotten similar results?

Heonozis commented 4 years ago

Unfortunately no. I was just testing the concept. First of all, I downscaled the network to use 64x64 images. And also, I did only 50 epochs (in the paper they did 200K). But conceptually it works and I'll use this technique for face swap.

Also, you can refer to the original repository to see the network for 256x256 input. It will take much more time to train, but the results will be significantly better. https://github.com/taotaonice/FaceShifter

Heonozis commented 4 years ago

Close to inactivity