World models, systems that synthesize realistic video sequences from an initial image and a set of actions, are becoming central to embodied AI, simulation, and robotics research. The core challenge has always been scaling these systems to generate minute-long, high-resolution video without demanding prohibitively large compute clusters. Most teams working in this space have had to pick a side: quality or accessibility. NVIDIA just made that tradeoff largely irrelevant.
SANA-WM is an efficient 2.6B-parameter open-source world model natively trained for one-minute generation, synthesizing high-fidelity, 720p, minute-scale videos with precise camera control. It's built on the NVlabs/Sana GitHub repository and documented on the official project page, with the full paper available on arXiv.
As someone who follows the video generation space closely, the efficiency numbers here are what stand out. This isn't incremental progress on an existing benchmark — it's a different resource profile entirely.
What Is SANA-WM?
SANA-WM takes a single image and a camera trajectory as input, then synthesizes a realistic 60-second, 720p video that faithfully follows that trajectory. Released in May 2026, SANA-WM supports 720p, 1-minute video generation with 6-DoF camera control and is positioned as a new baseline for world modeling and embodied AI.
Most competitive open-source baselines either require multi-GPU inference or sacrifice resolution to stay within compute budgets. NVIDIA's SANA-WM directly targets these bottlenecks. The model is built on the SANA-Video codebase, which itself was accepted as an oral presentation at ICLR 2026.
The Architecture That Makes It Work
Generating a 60-second video at 720p means modeling 961 latent frames. Standard softmax attention, the default in most video diffusion models, has memory and compute that grows quadratically with sequence length. At that scale, standard transformers simply don't fit in memory.
Four core designs drive the architecture: Hybrid Linear Attention combines frame-wise Gated DeltaNet (GDN) with softmax attention for memory-efficient long-context modeling, while Dual-Branch Camera Control ensures precise 6-DoF trajectory adherence.
SANA-WM replaces most attention blocks with frame-wise Gated DeltaNet (GDN). Unlike token-wise GDN used in language models, SANA-WM's frame-wise variant processes one entire latent frame per recurrent step. The GDN update rule keeps the recurrent state at a constant D×D size regardless of video length.
A dual-branch camera control module runs in parallel with the main generation trunk. One branch encodes the 6-DoF trajectory. The other handles content synthesis. They merge at each transformer block, giving the model continuous access to camera position data throughout the full generation pass.
The remaining two architectural pillars are equally important:
- Two-Stage Generation Pipeline applies a long-video refiner to stage-1 outputs, improving quality and consistency across sequences.
- Robust Annotation Pipeline extracts accurate metric-scale 6-DoF camera poses from public videos to yield high-quality, spatiotemporally consistent action labels.
Performance and Benchmarks
SANA-WM uses only approximately 213K public video clips with metric-scale pose supervision, completes training in 15 days on 64 H100s, and generates each 60-second clip on a single GPU. Its distilled variant can be deployed on a single RTX 5090 with NVFP4 quantization to denoise a 60-second 720p clip in 34 seconds.
