Back to Gap Map Model components / Image, video, 3D & music generation

OmniGen

VectorSpaceLab
restricted / Overall score: 1.6

OmniGen is VectorSpaceLab's line of all-in-one image generation models. OmniGen2, released in June 2025, handles text-to-image generation, instruction-based editing and in-context generation that places given subjects into a new scene, by pairing the Qwen2.5-VL vision-language model with a separate image decoder. The team also published its training datasets and an in-context generation benchmark.

OmniGen v1 (October 2024, MIT) sits on a separate Hub account, Shitao/OmniGen-v1, and is not counted here. OmniGen2's own transformer and code are Apache 2.0, but the checkpoint bundles Qwen2.5-VL-3B-Instruct under Alibaba's Qwen Research License, which allows research and evaluation only. The EditScore reward models from the same team are not part of this entry.

Openness

2 medium confidence
2.0
weights
open (the OmniGen2 checkpoint, including its bundled Qwen2.5-VL-3B component, downloads from the Hub without a gate)
data
partial (the X2I2 training datasets are released under Apache 2.0; the data construction pipeline is not yet)
code
partial (inference code plus full-parameter and LoRA fine-tuning scripts; the pretraining pipeline is not published)
license
Apache-2.0 (the OmniGen2 transformer weights and the repository code)+Qwen-Research-License (the Qwen2.5-VL-3B-Instruct model shipped in the checkpoint's mllm/ folder, which every OmniGen2 run loads: non-commercial use only)

OmniGen2's own weights and code are Apache 2.0 and its training datasets are published, but the checkpoint ships Alibaba's Qwen2.5-VL-3B as its understanding model, and that component is licensed for research and evaluation only. Since every OmniGen2 run needs it, the model as distributed cannot be used commercially.

Adoption

1 high confidence
1.0

Adoption is measured as Hugging Face downloads of the OmniGen2 and OmniGen2-RL checkpoints on the project's Hub account; the EditScore reward models there, and OmniGen v1 on a personal account, are left out.

Capability

2 medium confidence
2.0

OmniGen2 is near the bottom of the public image arena, which scores text-to-image only and not the editing and subject-driven generation the model was built for. FLUX.2 [dev] places well above it.

Verified 2026-11-12