Cosmos3-Edge checkpoint and inference recipe update — August 2026
#62
by jjf233 - opened
Update — August 25, 2026
The Cosmos3-Edge generator checkpoint, runtime defaults, usage examples, and benchmark results have been updated. If you pull from main, please refresh your local snapshot.
What changed
- Checkpoint: updated Cosmos3-Edge generator weights.
- Inference recipe (image-to-video): 480p generation at
832×480,121frames,20denoising steps, guidance scale6.0, flow shift12.0. - Usage examples: the Cosmos3-Edge cookbook section is now image-to-video only — the text-to-image and text-to-video Edge examples were removed from both the PyTorch/Diffusers and vLLM-Omni notebooks.
- Benchmarks: refreshed PAIBench-G, RBench, PhysicsIQ, and throughput results.
Updated benchmark results (image-to-video, 480p / 24 fps)
| Model | PAIBench-G Overall ↑ | Domain ↑ | Quality ↑ | RBench ↑ | PhysicsIQ ↑ | Throughput (frames/s) ↑ |
|---|---|---|---|---|---|---|
| Cosmos3-Edge | 81.1 | 84.2 | 78.1 | 55.1 | 34.1 | 6.71 |
| Cosmos3-Nano | 82.4 | 86.5 | 78.2 | 54.5 | 41.2 | 2.66 |
| Cosmos-Predict2.5-2B | 80.0 | 83.0 | 76.9 | 44.6 | — | 1.31 |
Throughput is frames generated per second, measured in eager mode on a single NVIDIA H100 GPU. Cosmos3-Edge delivers the highest generation throughput while staying competitive on quality.
How to refresh
huggingface-cli download nvidia/Cosmos3-Edge --local-dir ./Cosmos3-Edge
# or, if you track the repo with git:
git -C Cosmos3-Edge pull
Compatibility & reproducibility
- Reproduce the image-to-video numbers with the updated recipe: 480p, 20 steps, guidance 6.0, flow shift 12.0.
- The self-contained Edge image-to-video example lives in the audiovisual generator cookbook (
run_with_diffusers.ipynb/run_with_vllm_omni.ipynb).
Use this thread for questions or issues about the update.