dorarad/gansformer

Combines bipartite transformer architecture with multiplicative integration for efficient high-resolution image synthesis, enabling long-range spatial interactions with linear computational complexity. The model alternates information flow between latent variables and visual features to encourage compositional object representations, generalizing StyleGAN's region-based modulation approach. Supports both TensorFlow and PyTorch implementations with pretrained weights for resolutions up to 1024×1024, trained convergence in 5-7x fewer steps than StyleGAN2.

1,346 stars. No commits in the last 6 months.