whatmodel.sh / image / OmniGen2
OmniGen2
↗ share
+ contribute
OmniGen2 is an open-weights image model from BAAI, released June 2025. Its published size is 4B parameters. Licence: apache-2.0.
Model information sources
Release date ↗ · Weights licence ↗ · Parameter count ↗
Diffusion decoder only (approx. 4B, the trainable part). The makers paper writes the model size as "3B + 4B": a frozen Qwen2.5-VL-3B-Instruct MLLM plus the 4B decoder; VAE excluded. Reviewer may prefer recording the 3B MLLM separately; no single total is stated.
Score within the LMArena text-to-image comparison set only.
$ how it compares
The image models ranked either side of OmniGen2, out of 197 with a score, then more from BAAI.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 134 |  Midjourney V6.1 | 42±4 | comparison only | ||
| 135 |  FLUX.1 [schnell] | 42±2 | api $0.0030/image +local | ||
| 136 | OmniGen2 | 42±2 | api $0.15/image +local | ||
| 137 |  Stable Diffusion 3.5 Large | 42±1 | api $0.065/image +local | ||
| 138 |  Show-o2 7B | 42±3 | local | ||
| more from BAAI | |||||
| 48 | Emu3.5-Image | 60±5 | local | ||
| 69 | Emu3.5 | 57±5 | local | ||
| 164 | Emu3 | 36±2 | local | ||
$ what it's good at
text to image▂▅▅▃▄▄█▃▁▂▁▁ #91 of 126 * · 21% of its score · 3 tests
Beats Latent Diffusion text2img-large (LDM-KL-8-G, LAION-400M), VQ-Diffusion (ITHQ) here, though they rank higher overall. Trails Show-o2 1.5B-HQ, Stable Diffusion 3.5 Medium, Playground v2.5 here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| ImagenWorld text-guided generation (human overall) | 74.0% ✎ | ▄··█▄·▄··▄·▄ beats 1/6 | 16% |
| XTC-Bench generation (Overall Gen., scene-graph node agreement) | 49.9% ✎ | ▄█·▄·······▄ beats 1/4 | 5.9% |
| OneIG-Bench (EN) lab | 47.0% ✎ | ▃······▃▅█▅▃ beats 0/9 | — |
image to image / editing▁▁·▃▂▃█▃▃▄▁▁ #55 of 62 · 57% of its score · 8 tests
Beats UNO, BAGEL here, though they rank higher overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Edit-Compass (EN, overall) | 37.6% ✎ | ▃▃▂·▃▆▂█·▃·▃ beats 7/20 | 22% |
| UnicBench (EN, overall) | 61.2% ✎ | ▂····▂▄▂▄·█▂ beats 2/11 | 10% |
| GEditBench v2 (overall Elo) | 888.0 ✎ | ▃▃····▂█▂▃▂▃ beats 2/14 | 6.7% |
| CREval-Bench creative image editing (Overall, GPT-4o QA evaluator) | 68.2% ✎ | ▄█▄··█··▄·▄▄ beats 1/8 | 5.9% |
| MultiTextEdit cross-lingual text-in-image editing (SemAvg, LVM judge) | 34.7% ✎ | ▄·▄▄▄▄··█▄·▄ beats 1/8 | 4.9% |
| ImagenWorld text-guided editing (human overall) | 58.0% ✎ | █··█··██···█ beats 1/4 | 4.3% |
| OmniContext average (GPT-4.1 judge) lab | 79.5% ✎ | ▃▃▃··▃·····█ beats 6/6 | 2.8% |
| GEdit-Bench-EN overall (G_O, GPT-4.1 judge) lab | 72.1% ✎ | ▄·····█·▄▄·▄ beats 4/5 | — |
other tests4 not in a task · 20% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| RealPref-Bench image editing (human user-study score) | 68.7% ✎ | █▄▄····▄···▄ beats 2/5 | 11% |
| WEAVEBench in-context multi-turn editing (Avg) | 60.9% ✎ | █▄▄·▄▄▄▄█·▄█ beats 9/12 | 6.3% |
| ViGoR-Bench zero-shot visual reasoning, image models (Result Avg) | 22.4% ✎ | ▂···▂·▃█▆▅·▂ beats 2/15 | 2.1% |
| Knowledge Visualization T2I (Brief Caption, English overall) | 18.5% ✎ | ▅·▃·▃··▃·█▃▃ beats 1/9 | 1.2% |
$ where to run it
Through an API
api · fal.ai · $0.15/image · fal.ai ↗ · price sourceUSD for one 1-megapixel output: fal bills fal-ai/omnigen-v2 (OmniGen V2 / OmniGen2) $0.15 per megapixel (price = rate x 1 MP; fal's rounding for 1024x1024 ~1.05 MP is not stated; larger outputs cost proportionally more). Product: fal.ai serverless image generation and editing. API model: fal-ai/omnigen-v2. Settings: 1 MP output, one image. Price basis: published per-megapixel list rate x 1 MP. Currency: USD. Not a BAAI direct tariff. fal.ai official model page, observed 2026-09-26.
Other ways to run it
- Run it as published · needs ~17 GB, NVIDIA only · from github.com/VectorSpaceLab
git clone https://github.com/VectorSpaceLab/OmniGen2.git && cd OmniGen2 && pip install torch==2.6.0 torchvision --extra-index-url https://download.pytorch.org/whl/cu124 && pip install -r requirements.txt && bash example_t2i.sh
Weights and documentation: huggingface.co/OmniGen2 ↗
$ how the score was worked out · Edit-Compass & EditReward-Com… 22%, ImagenWorld: Stress-Testing I… 20%, HP-Edit: A Human-Preference P… 11% and 9 more
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
not counted (2)
| BAAI — OmniGen2 technical report | GEdit-Bench-EN overall (G_O, GPT-4.1 judge) [correct] | 72.1% | fewer than three models have this test yet, or this source ran it for this model alone |
| BAAI — OmniGen2 technical report | OneIG-Bench (EN) [correct] | 47.0% | fewer than three models have this test yet, or this source ran it for this model alone |
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json