open-mmlab/PowerPaint

Built on a diffusion architecture with task-specific prompt embeddings, PowerPaint uses tailored text prompts (e.g., "P_obj" for insertion, "P_ctxt" for removal) routed through cross-attention layers to unify four distinct inpainting tasks in a single model. The v2 iteration integrates BrushNet-style conditioning while preserving cross-attention mechanisms for task prompt injection, and supports ControlNet modules (Canny, Depth, Pose) for additional spatial control during generation. Distributed via Hugging Face and OpenXLab with Gradio-based inference, the project integrates into the MMagic ecosystem and provides both model weights and inference code.

1,059 stars.