OmniGen2

OmniGen2 folds four capabilities into one open model: visual understanding, text-to-image generation, instruction-guided image editing, and in-context generation that composes new scenes from reference images of people and objects. Where the original OmniGen used a single shared pathway, this successor from VectorSpaceLab at BAAI runs separate decoding pathways for text and image with unshared parameters and a decoupled image tokenizer, preserving language ability while improving generation quality. Released in June 2025 with a technical report on arXiv, it reaches state-of-the-art instruction-editing results among open models according to the authors. Deployment needs an NVIDIA GPU with roughly 17 GB of VRAM, an RTX 3090 being the reference card, and CPU offload roughly halves that at a small speed cost; installation is a Python 3.11 environment with PyTorch 2.6 and optional flash-attn. TeaCache and TaylorSeer acceleration are supported, training and fine-tuning code is public, and both the code and the weights carry the Apache-2.0 license, so commercial use is unrestricted. The repository holds about 4,100 stars.