whatmodel.sh / image / Cosmos3-Super-Text2Image

Cosmos3-Super-Text2Image

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Cosmos3-Super-Text2Image is an open-weights image model from NVIDIA, released May 2026. Its published size is 64B parameters. Licence: openmdw-1.1.

62.8±2.3#29 of 144 image modelsfrom 5 benchmark results across 2 sources

Score within the LMArena text-to-image comparison set only.

$ how it compares

The image models ranked either side of Cosmos3-Super-Text2Image, out of 197 with a score, then more from NVIDIA.

#modelscorewhereRAMreleased
33Wan 2.7 Image Pro Alibaba63±2api $0.075/image2026-04
34Reve V1.5 Reve63±3comparison only2026-02
35Cosmos3-Super-Text2Image NVIDIA63±2local~154GB2026-05
36FLUX.2 [pro] Black Forest Labs63±2api $0.030/image2025-11
37FLUX.2 [flex] Black Forest Labs62±2api $0.050/image2025-11
more from NVIDIA
130ChronoEdit-14B NVIDIA43±3local~34GB2025-10

see all image models · find the ones that fit your machine

$ what it's good at

text to image▂▅▅▃▄▄█▃▁▂▁▁ #31 of 126 · 88% of its score · 4 tests

Beats Wan 2.7 Image Pro here, though they rank higher overall. Trails ERNIE-Image, Nano Banana (Gemini 2.5 Flash Image), FLUX.2 [flex] here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
LMArena text-to-image1151.6 ✎▂▂▃▄▇█▄▂▄▂·▂ beats 46/75avg 1027▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 82%typical
HPSv3 score of Cosmos 3 report text-to-image evaluation images lab11.6000 ✎▄···█··▄·▄·▄ beats 3/5avg 11.076▇▇ 5.1%blunt
CVTG-102ch PNED (Chinese character rendering, Cosmos 3 report) lab41.2% ✎▄·······▄█▄▄ beats 0/5medium · avg 60%▇ 1.3%blunt · at its level · noisy
CVTG-500L PNED (English long-prompt text rendering, Cosmos 3 report) lab89.1% ✎█······█████ beats 4/5easy · avg 85%▇ 0.0%blunt
other tests1 not in a task · 10% of its score
testresultvs the others in its rundifficultycountswhy
UniGenBench + Physical-AI prompts, All 1,170 (Cosmos 3 report) lab91.4% ✎▄···█··▄···█ beats 5/5easy · avg 78%▇▇▇▇ 10%typical

$ where to run it

Other ways to run it

  1. Run it as published · needs ~154.1 GB, NVIDIA only · from huggingface.co/nvidia (official)
    uv run --with git+https://github.com/huggingface/diffusers --with transformers --with accelerate --with torch --with torchvision --with av --with imageio --with imageio-ffmpeg --with cosmos_guardrail python -c "import torch; from diffusers import Cosmos3OmniPipeline; from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler; p=Cosmos3OmniPipeline.from_pretrained('nvidia/Cosmos3-Super-Text2Image', torch_dtype=torch.bfloat16, device_map='cuda', enable_safety_checker=True); p.scheduler=UniPCMultistepScheduler.from_config(p.scheduler.config, flow_shift=3.0); p(prompt='a cat holding a sign', negative_prompt='', num_frames=1, height=1024, width=1024, num_inference_steps=50, guidance_scale=4.0).video[0].save('out.png')"

Official weights and documentation: huggingface.co/nvidia ↗

$ how the score was worked out · LMArena leaderboard dataset 82%, Cosmos 3: Omnimodal World Mod… 16%

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

sourceresultsshare of the scorelicence
LMArena leaderboard dataset1▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 82%CC BY 4.0
Cosmos 3: Omnimodal World Models for Physical AI (arXiv:2606.02800v4) · lab-reported4▇▇▇▇▇ 16%CC-BY-4.0
prior · pull toward the middle▇ 2%
weighted average= 62.8

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

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