NVIDIA's SANA-WM: 2.6B Open-Source World Model Generates Minute-Long 720p Video on One H100
NVIDIA Labs has released SANA-WM, a 2.6-billion-parameter open-source world model that takes a single image and a 6-DoF camera path and produces a full minute of 720p video. The model runs at inference on a single H100. Training required 64 H100s for 15 days.
The paper (arxiv 2605.15178) drops as AI video generation splits into two categories: general-purpose text-to-video, and world models built around physical camera control and scene coherence over long horizons. SANA-WM is firmly in the second category, and it gets there at a fraction of the compute of industrial alternatives like LingBot-World and HY-WorldPlay.
The Architecture
Four components drive SANA-WM’s efficiency:
Hybrid Linear Attention pairs frame-wise Gated DeltaNet with periodic softmax attention. The DeltaNet handles per-frame temporal context cheaply; the softmax passes maintain long-range coherence across the full minute. The combination avoids the memory blow-up that forces most long-video models to limit horizon length.
Dual-Branch Camera Control splits trajectory adherence into a coarse global pose branch and a fine pixel-aligned geometric branch. Together they track metric-scale 6-DoF camera paths with precision that single-branch approaches miss at longer sequences.
Two-Stage Generation Pipeline uses the 2.6B backbone for temporal structure and dynamics, then applies a 17B long-video refiner to sharpen texture and correct degradation in late windows. NVIDIA calls this “Bet Small to Win Big”: you pay the large-model cost only on a second pass over latent representations, not at generation scale.
Robust Annotation Pipeline extracts accurate metric-scale 6-DoF camera poses from public video data, producing spatiotemporally consistent action labels for training.
Numbers
On VBench at 720p, SANA-WM scores 84.05 total (84.63 quality, 81.73 semantic alignment). Against the field:
| Model | Params | Latency (s) | VBench Total |
|---|---|---|---|
| Wan-2.1-14B | 14B | 1,897 | 83.73 |
| Wan-2.1-1.3B | 1.3B | 400 | 83.38 |
| Wan-2.2-5B | 5B | 116 | 83.28 |
| SANA-WM | 2.6B | 36 | 84.05 |
36 seconds against 1,897 for Wan’s flagship. The 2B model outscores the 14B on total and semantic while running 52x faster.
That latency gap matters practically. Running Wan-2.1-14B to generate one minute of test video takes over 30 minutes of wall time. SANA-WM generates the same clip before a browser tab finishes loading.
Context
SANA-WM builds on the SANA family, which has been solving the compute-per-quality problem across modalities since 2024. The SANA-Video predecessor was an ICLR 2026 Oral. The 720p model with LTX-VAE landed in March 2026.
The WM variant is the first in the family designed explicitly for world modeling: controllable camera paths, scene coherence over 60+ seconds, and output suitable for training downstream robotics or simulation policies, not just content generation.
Code is under Apache 2.0. Weights are on HuggingFace under the NVIDIA Open Model License. The 2K refiner path uses LTX-2 as the second stage, available separately.
Why It Matters
World models are the infrastructure play behind physically capable AI. Robotics labs, autonomous vehicle teams, and simulation-first training pipelines all need the same thing: controllable, coherent long-horizon video generated cheaply enough to produce at dataset scale.
SANA-WM is the first open model to deliver minute-scale world rollouts on a single consumer-grade GPU at quality that matches closed industrial systems. The training cost of 64 H100-days is low enough that academic labs and mid-size companies can run their own fine-tuning passes.
That’s what changes the timeline: it’s not that NVIDIA shipped a better video model, it’s that they made world modeling accessible outside the lab.