OmniGen2

OmniGen2 is an open-source unified multimodal model from VectorSpaceLab, released June 16, 2025. Unlike the original OmniGen, it uses two distinct decoding pathways for text and image modalities with unshared parameters and a decoupled image tokenizer. Its architecture pairs a 3B vision-language model (built on Qwen2.5-VL) with a 4B diffusion model, covering four capabilities: visual understanding, text-to-image generation, instruction-guided image editing (state-of-the-art among open models at release), and in-context generation — flexibly combining reference humans, objects, and scenes into novel coherent outputs. Weights are available on Hugging Face and ModelScope, with Gradio demos published.

The honest picture: at 7B total parameters it is efficient but cannot match the raw fidelity of much larger closed systems like GPT-4o image generation or Imagen on complex scenes. The team promised training code, datasets, and a ComfyUI integration that were still on the TODO list as of release — check the repo for current status. In-context generation is powerful but finicky; multi-reference compositions need careful prompt engineering to stay coherent.

Clone the VectorSpaceLab/OmniGen2 repo or load weights from Hugging Face (OmniGen2/OmniGen2). Gradio demos available for quick tries.