lama

Why we included this project

LaMa is the reference implementation behind several popular object removal and cleanup tools, so teams building photo editing features will recognize it as a proven core. It fills arbitrary masked regions of an image, such as a person, cable, or watermark, using a GAN built on Fourier convolutions, which the authors show handles large masks and periodic textures better than most earlier approaches. The resolution-robust behavior is the real draw: it trained on small crops yet generalizes to roughly 2K input, so you can slot it into an editing pipeline without retraining for every image size. Pretrained weights for Places2 and CelebA-HQ are available, and you can run inference from the command line or a Docker image, which makes evaluating it against your own images quick. The project is well documented and has spawned several third-party wrappers, a useful on-ramp if you want something lighter than the raw research code.