ali-vilab/AnyDoor

Built on Stable Diffusion v2.1 and ControlNet, AnyDoor leverages DINOv2 for semantic feature extraction to enable precise object-level customization without task-specific training. The system accepts reference objects and target masks, using a diffusion-based architecture to synthesize realistic placements while preserving object identity and scene context. Training uses multi-dataset supervision (COCO, UVO, LVIS) and supports inference on downstream applications like virtual try-on and face swapping through optional domain-specific fine-tuning.

4,222 stars. No commits in the last 6 months.