SANA-WM, an Open-Source 2.6B-Parameter World Model for 1-Minute 720p Video

NVIDIA has unveiled SANA-WM. It is a 2.6B-parameter world model that takes a single image and a 6-DoF camera trajectory as input and generates controllable 720p, 1-minute-long video on a single GPU. The page provides Paper, Code, and Models soon links.

The core architecture is a Hybrid Linear Diffusion Transformer.

• It combines frame-wise Gated DeltaNet with periodic softmax to maintain consistency over long rollouts.
• A coarse global pose branch and a fine pixel-aligned geometry branch improve camera path following.
• The all-softmax variant is described as running into OOM during 60-second generation.

Training used about 213K publicly available videos with meter-scale 6-DoF pose supervision.

• Full training took 15 days on 64 H100s.
• Inference is possible on a single H100.

A 17B long-video refiner is attached for quality improvement.

• It refines the texture, motion, and later-segment quality of the stage-1 output.
• The distilled variant denoises a 60-second 720p clip in 34 seconds using NVFP4 on 1 RTX 5090.

On benchmarks, it reportedly showed higher action-follow accuracy than existing open-source baselines, and 36x higher throughput at comparable visual quality.