AttentionGAN

Why we included this project

AttentionGAN is the official PyTorch implementation of a method for unpaired image-to-image translation that adds learned foreground and background attention masks to a CycleGAN-style generator. Rather than translating the whole image at once, the model learns which regions of the output to keep and which parts of the input background to preserve, which shows up clearly on tasks like selfie-to-anime, horse-to-zebra, and map-to-aerial-photo conversion. The repo includes pretrained models, dataset download scripts, and training and test commands, so you can reproduce the paper results or adapt the code to your own domain pairs. There is also a multi-domain variant for things like facial attribute transfer and a geometry-aware version for translations that involve shape changes. One thing to watch: the code is released under a non-commercial license for academic research, so teams planning commercial deployment should contact the authors first.