search

boyzwj / zface

improve from hififace
27 stars 3 forks source link

Have you compared the results before and after modified? #2

Open VitoChien opened 2 years ago

VitoChien commented 2 years ago

Hi, thank you for sharing the code! I noticed that you changed the AdaIN to modulated convolution and the discriminator is also changed to which released by SimSwap. I wonder if you have compared the results before and after modified?

boyzwj commented 2 years ago

The generator network from hififace or simswap both has the problem of “Checkerboard Artifact” on my training results，the longer I trained，the problem more obvious，as we knew the stylegan2 solves this problem very well，I used the modulated convolution AdaIN layer reference from stylegan2，the problem was gone，It give me a very high quality result.. The discriminator referenced to stylegan-xl ,used 2 multiscale pretrained feature network , compare to origin multiscale disciminator (from stargan-v2) , I got a better result in the same training time, obviously the pretrained network has much more parameters and we don't need extral memory and computing power to train.

zhanghm1995 commented 2 years ago

@boyzwj Hi, thank you for sharing this code!

I want to know whether the "Checkerboard Artifact" you mentioned is like below (first row: source image, second row: target image, last row: generated faces): image

When I use the hififace implemented by this one https://github.com/mindslab-ai/hififace and trained it in my own dataset, I always cannot get promising results. There are so many artifacts.