EndlessSora/DeceiveD

Implements Adaptive Pseudo Augmentation (APA), a discriminator-deception strategy that leverages generator-synthesized images to augment real data distributions during training, directly addressing discriminator overfitting in low-data regimes. Built on StyleGAN2-ADA-PyTorch architecture with seamless integration and minimal computational overhead, supporting multi-GPU training across varied datasets (FFHQ, AFHQ, anime, birds) at resolutions up to 1024×1024. Provides pretrained models, dataset preparation utilities, and configurable training pipelines with FID evaluation against full reference datasets.

235 stars. No commits in the last 6 months.