search

zj-dong / AG3D

Official code release for ICCV2023 paper AG3D: Learning to Generate 3D Avatars from 2D Image Collections
https://zj-dong.github.io/AG3D/
245 stars 21 forks source link

Is there a problem with the output of the discriminator? #14

Open ripperthejack opened 9 months ago

ripperthejack commented 9 months ago

The generator is pre-trained and I used the dual discriminator just like:

ag3d_discriminator = legacy.load_network_pkl(f)['D'].to(args.device) output = G.only_synthesis(ws=ws, c=c, truncation_psi=truncation) full_logit = ag3d_discriminator.D_image(output,c)

but i got a low confidence around -10 , is that normal?

shash29-dev commented 7 months ago

Hi, There is something it seems regarding discriminators. I tried -- Training from Scratch: fake_depths and fake_images look alright (similar to sampled smpl pose). As training proceeds, fake_depths become total black and white. images go full background-- (white) Training from provided weights: fake_depths and fake_images look alright (similar to sampled smpl pose, and realistic). As training proceeds, both deteriorate towards ---> fake_depths become total black and white. images go full background-- (white)

I speculate the reason behind it, may be a discriminator issue driving the generator crazy?

ripperthejack commented 7 months ago

hey bro , have you successfully reproduced this model ?