advimman/lama

Employs Fourier convolutions in its encoder-decoder architecture to achieve resolution robustness, enabling inpainting at ~2K resolution despite training on 256x256 images, with particular strength on periodic structures. Integrates with PyTorch and PyTorch Lightning, and provides pre-trained models for Places and CelebA-HQ datasets with optional feature refinement post-processing. Widely adopted across third-party implementations including Hugging Face Spaces with Gradio, desktop applications (PyQt5), mobile formats (Core ML), and composable detection+inpainting pipelines (DETR, MaskDINO).

9,765 stars. No commits in the last 6 months.