NVIDIA · 2026-05-14 · major
SANA-WM — NVIDIA's 2.6B Open-Source World Model Generates 720p One-Minute Video With 6-DoF Camera Control
A 2.6B-parameter open-source world model that synthesizes 720p, minute-long video with precise 6-degree-of-freedom camera control, using hybrid linear attention to keep long-context generation memory-efficient.
An efficient open-source world model that generates minute-long, camera-controllable 720p video.
Key specs
| Parameters | 2.6B |
|---|---|
| Resolution | 720p |
| Max clip length | 60s |
| Throughput | 36x vs prior open-source |
| Training data | ~213K video clips |
What is it?
SANA-WM is a 2.6-billion-parameter world model from NVIDIA that generates high-fidelity 720p video up to one minute long. Unlike a plain text-to-video model, it accepts 6-degree-of-freedom camera trajectories as input, so the virtual camera can be steered through the generated scene. It ships as open source, positioned as a baseline for world modeling and embodied AI research.
How does it work?
The model is a hybrid linear diffusion transformer. Frame-wise Gated DeltaNet linear attention is combined with softmax attention to model long video context without the memory cost of full attention. A dual-branch camera-control module enforces 6-DoF trajectory adherence, and a two-stage pipeline runs a long-video refiner to improve consistency. It was trained on roughly 213K public video clips with metric-scale pose annotations.
Why does it matter?
Minute-scale, camera-controllable video generation has mostly required large industrial systems. SANA-WM reports comparable visual quality at just 2.6B parameters with a claimed 36x throughput gain over prior open-source world models, putting this line of research within reach of smaller labs and embodied-AI teams.
Who is it for?
world-model and embodied-AI researchers
Try it
https://github.com/NVlabs/Sana