Projected GANs Converge Faster

Generative Adversarial Networks (GANs) produce high-quality images but are challenging to train. They need careful regularization, vast amounts of compute, and expensive hyper-parameter sweeps... We make significant headway on these issues by projecting generated and real samples into a fixed, pretrained feature space. Motivated by the finding that the discriminator cannot fully exploit features from deeper layers of the pretrained model, we propose a more effective strategy that mixes features across channels and resolutions. Our Projected GAN improves image quality, sample efficiency, and convergence speed. It is further compatible with resolutions of up to one Megapixel and advances the state-of-the-art Fr\'echet Inception Distance (FID) on twenty-two benchmark datasets. Importantly, Projected GANs match the previously lowest FIDs up to 40 times faster, cutting the wall-clock time from 5 days to less than 3 hours given the same computational resources. read more

PDF Abstract NeurIPS 2021 PDF NeurIPS 2021 Abstract
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Image Generation ADE-Indoor Projected GAN FID 6.7 # 1
Image Generation Cityscapes Projected GAN FID-10k-training-steps 3.41 # 1
Image Generation CLEVR Projected GAN FID-5k-training-steps 0.89 # 1
Image Generation CUB 128 x 128 Projected GAN FID 2.79 # 1
Image Generation FFHQ 256 x 256 Projected GAN FID 2.2 # 1
Image Generation LSUN Bedroom 256 x 256 Projected GAN FID 1.52 # 1
FID-10k-training-steps 1.52 # 1
Image Generation LSUN Cat 256 x 256 Projected GAN FID 3.89 # 1
Image Generation LSUN Churches 256 x 256 Projected GAN FID 1.59 # 1
Image Generation LSUN Horse 256 x 256 Projected GAN FID 2.17 # 1
Image Generation Oxford 102 Flowers 256 x 256 Projected GAN FID 3.86 # 1
Image Generation Stanford Cars Projected GANs FID 2.09 # 1
Image Generation Stanford Dogs Projected GAN FID 11.75 # 1

Methods


No methods listed for this paper. Add relevant methods here