clementabary / slegan

PyTorch (Lightning) implementation of SLE-GAN

Few-shot Adversarial Image Synthesis with SLE-GAN

PyTorch Lightning-based implementation of ICLR2021's "Towards faster and stabilized GAN training for high-fidelity few-shot image synthesis" (unofficial).

The generator needs to compare favorably to StyleGAN2 with latest model configuration and differentiable data augmentation for best few-shot training performance.

Build a strong baseline for G :

• spectral normalization (over D or G)
• exponential-moving-average optimization on G
• differentiable augmentation (over D)
• GLU instead of ReLU in G

Add the two proposed techniques :

• skip-layer excitation module
• self-supervised discriminator

Refinements (not in paper but in official code) :

• LPIPS-VGG perceptual loss for reconstruction
• Label smoothing in hinge loss
• Noise injection layer
• Swish activation in SLE blocks
• Auxiliary 128-sized layer output

Miscellaneous to-do :

• Add FID tracking (every 10k iterations)
• Add sampling with truncation
• Add interpolation tools
• Add style mixing pipeline

First samples

The following image grid of size 1024 has been generated at the 65k-th iteration (on a "one day - one P100" basis) with the main configuration as is.