JIA-Lab-research/DreamOmni2

Leverages a unified diffusion-based architecture with separate LoRA modules for editing and generation tasks, using multimodal encoders to process both text instructions and reference images for concrete object or abstract attribute guidance. Supports both subject-driven generation with identity/pose consistency and inpainting-aware editing that preserves non-edited regions while accepting visual references alongside natural language prompts. Available on Hugging Face with web demo interfaces and integrated with ComfyUI for production workflows.

2,273 stars.