gansformer

Why we included this project

Researchers and practitioners in image synthesis who want a transformer-based alternative to convolutional GANs will find this repository useful. It includes both TensorFlow and PyTorch implementations of GANformer, with pre-trained checkpoints for 256×256 scene datasets and outputs up to 1024×1024 for FFHQ and Cityscapes. What sets the model apart is its bipartite attention structure: information flows through a small set of latent variables rather than every pairwise interaction, so compute stays roughly linear and high-resolution synthesis does not blow up memory. The generate.py script samples from a downloaded model in about ten lines, and a truncation-psi knob offers a direct quality-versus-diversity trade-off, making it easy to reproduce the ICML paper's results or compare attention-based generators against StyleGAN-style baselines. It is also a sensible entry point if you are curious how transformer ideas generalize from language to visual generation.