wpeebles/gangealing

Leverages joint end-to-end training of a Spatial Transformer network and StyleGAN2 generator, where the transformer warps unaligned GAN samples to a learned canonical template while the generator adapts to minimize alignment difficulty. Pre-trained models enable zero-shot dense correspondence and object propagation on real images despite training exclusively on synthetic data, with applications ranging from video tracking to mixed reality effects. Includes optimized CUDA kernels for anti-aliased grid sampling and splatting visualization, plus distributed training support via PyTorch's torchrun API.

1,013 stars. No commits in the last 6 months.