search

jgkwak95 / SURF-GAN

[ECCV 2022] Official Pytorch implementation of "Injecting 3D Perception of Controllable NeRF-GAN into StyleGAN for Editable Portrait Image Synthesis"
MIT License
126 stars 14 forks source link

Issues reproducing results #10

Closed NOlivier-H closed 1 year ago

NOlivier-H commented 1 year ago

Hello, thanks for your work, and for making your code available. I am having issues reproducing your results on CelebA, the network seems to converge up to around iteration 15k, but then the image quality degrades until becoming nothing recognizable. I used the command python train_surf.py --output_dir exp_name --curriculum CelebA_single, with nothing changed from your code. The python env was created using your instructions in "Create virtual environment". I ran the training process 3 times, and consistently obtained this result. Here are random samples at 15k, and 40k: 15k 40k

jgkwak95 commented 1 year ago

Hi @NOlivier-H , Thanks for you attention.

I just checked the train and curriculum files, I found that the batch size (now 14) is wrong. For the first stage, I used the batch_size: 56 (not 14) At each stage, the batch size is multiplied by 1/4. I forgot to change the original value after experimenting with my code under different conditions.

Therefore, please change the value of batch_size of the first stage in curriculums.py-CelebA_single. Generally, with higher batch_size, you can expect higher quality. so set the maximum batch size for your GPU.

jgkwak95 commented 1 year ago

I changed batch_size of the first stage 14-->56