A StyleGAN3 model trained on the danbooru2019 portraits dataset using vision-aided GAN. Sponsored by RunPod
I trained using vanilla StyleGAN3 up to 900 kimg, at which point I switched to vision-aided GAN for the remainder of training.
As I trained using vision-aided GAN, you cannot use this model with vanilla StyleGAN3 without some modifications, I made these modifications already here. You could also use vision-aided GAN's StyleGAN3
Used with RunPod's Pytorch pod. ipynb
Notes:
- s/kimg & GPU mem are with vision-aided GAN.
| Config | s/kimg (A6000) | GPU mem | Options |
|---|---|---|---|
| StyleGAN3-T | 54.84 | 11.50 | --cfg=stylegan3-t --gpus=4 --batch=32 --gamma=6 --dlr=0.001 --glr=0.001 --aug=noaug |
Note: If you are going to resume training you need to use these parameters or you will get a module not found error.
--cv=input-clip+dino+vgg+face_normals+face_seg-output-conv_multi_level+conv_multi_level+conv+conv+conv \
--cv-loss=multilevel_sigmoid_s+multilevel_sigmoid_s+sigmoid_s+sigmoid_s+sigmoid_s \
--warmup=0
Gwern Branwen, Anonymous, & The Danbooru Community; “Danbooru2019 Portraits: A Large-Scale Anime Head Illustration Dataset”, 2019-03-12. Web. Accessed 2022-06-19 https://www.gwern.net/Crops#danbooru2019-portraits
An anonymous friend who helped me with configuration and showed me vision-aided GAN