open-mmlab/PIA

Implements plug-and-play motion modules injected into Stable Diffusion's UNet to decouple motion generation from content preservation, enabling fine-grained control over motion magnitude while maintaining image fidelity. Integrates with DreamBooth-LoRA for personalized style adaptation and leverages PyTorch 2.0's scaled dot-product attention for memory-efficient inference on consumer GPUs (16GB VRAM for 1024×1024 images). Supports multi-framework deployment via HuggingFace, Replicate, and OpenXLab with configurable YAML-based inference pipelines.

978 stars. No commits in the last 6 months.