
摘要
尽管在生成图像建模方面取得了近期进展,但从复杂数据集(如ImageNet)中成功生成高分辨率、多样化的样本仍然是一个难以实现的目标。为此,我们尝试了迄今为止最大规模的生成对抗网络(Generative Adversarial Networks, GANs)训练,并研究了特定于这种规模的不稳定性。我们发现,对生成器应用正交正则化可以使其适用于简单的“截断技巧”(truncation trick),通过减少生成器输入的方差来精细控制样本保真度和多样性之间的权衡。我们的修改导致了在类别条件图像合成方面达到新的技术水平的模型。当在128x128分辨率的ImageNet上进行训练时,我们的模型(BigGANs)达到了166.5的Inception分数(IS)和7.4的Frechet Inception距离(FID),相比之前最佳的IS 52.52和FID 18.6有了显著提升。
代码仓库
abhi8585/GeneratedART-NFT-VISION
tf
GitHub 中提及
yaxingwang/MineGAN
tf
GitHub 中提及
rkorzeniowski/bigbigan-pytorch
pytorch
GitHub 中提及
yaxingwang/DeepI2I
pytorch
GitHub 中提及
clu0/unet.cu
pytorch
GitHub 中提及
mingtaoguo/biggan-tensorflow
tf
GitHub 中提及
vincent-thevenin/Realistic-Neural-Talking-Head-Models
pytorch
GitHub 中提及
minyoungg/GAN-Transform-and-Project
pytorch
GitHub 中提及
guillaumeboniface/super_resolution
tf
GitHub 中提及
alexverine/precisionrecallbiggan
pytorch
GitHub 中提及
ZVK/Talking-Heads
pytorch
GitHub 中提及
ZVK/talking_heads
pytorch
GitHub 中提及
ivclab/BigGAN-Generator-Pretrained-Pytorch
pytorch
GitHub 中提及
taki0112/BigGAN-Tensorflow
tf
GitHub 中提及
harpreetsodhi/ChangeMyPet-Deep-Learning-Model
pytorch
GitHub 中提及
ANIME305/Anime-GAN
tf
GitHub 中提及
kidist-amde/biggan-pytorch
pytorch
GitHub 中提及
ANIME305/Anime-GAN-tensorflow
tf
GitHub 中提及
gcervantes8/Game-Image-Generator
pytorch
GitHub 中提及
AlexVerine/PrecisionRecallGan
pytorch
GitHub 中提及
luqui/big-sleep
pytorch
GitHub 中提及
minyoungg/pix2latent
pytorch
GitHub 中提及
harpreetsodhi/ChangeMyPet_Deep_Learning_Model
pytorch
GitHub 中提及
sagy101/SoundGAN
pytorch
GitHub 中提及
yuanyuan-yuan/neural-coverage
pytorch
GitHub 中提及
rajveen/ChangeMyPet
pytorch
GitHub 中提及
Mohanned-Elkholy/BigGan-model-inversion-with-image-transformations
pytorch
GitHub 中提及
open-mmlab/mmgeneration
pytorch
uoguelph-mlrg/instance_selection_for_gans
pytorch
GitHub 中提及
ajbrock/BigGAN-PyTorch
pytorch
GitHub 中提及
lucidrains/big-sleep
pytorch
GitHub 中提及
amanjaiswal73892/change_my_pet
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| conditional-image-generation-on-artbench-10 | BigGAN + DiffAug | FID: 4.055 |
| conditional-image-generation-on-cifar-10 | BigGAN | FID: 14.73 Inception score: 9.22 |
| conditional-image-generation-on-imagenet | BigGAN-deep | FID: 5.7 Inception score: 124.5 |
| conditional-image-generation-on-imagenet | BigGAN | FID: 8.7 Inception score: 98.8 |
| image-generation-on-imagenet-128x128 | BigGAN | FID: 8.7 IS: 98.8 |
| image-generation-on-imagenet-128x128 | BigGAN-deep | FID: 5.7 IS: 124.5 |
| image-generation-on-imagenet-256x256 | BigGAN-deep | FID: 8.1 |