SANA-WM, a 2.6B open-source world model for 1-minute 720p video

SANA-WM is a 2.6B open-source world model released by NVIDIA that generates 60-second, 720p controllable video from just a single image and a camera trajectory, based on a Hybrid Linear Diffusion Transformer architecture.

The core structure consists of four elements.

• Hybrid Linear Attention: combines frame-wise Gated DeltaNet with periodic softmax to maintain long context.
• Dual-Branch Camera Control: follows 6-DoF paths with high precision using a coarse global pose branch and a fine pixel-aligned geometric branch.
• Two-Stage Generation Pipeline: after a stage-1 rollout, a 17B long-video refiner enhances texture, motion, and later-segment consistency.
• Robust Annotation Pipeline: extracts metric-scale 6-DoF pose from public videos to produce well-aligned action labels.

Training used approximately 213K public video clips, completed in 15 days on 64 H100s. The original model generates a 60-second 720p clip on a single H100, and the distilled variant denoises a clip of the same length in 34 seconds on an RTX 5090 using NVFP4 quantization.

On its own one-minute world-model benchmark, it reportedly shows higher action-following accuracy than existing open-source baselines, while maintaining visual quality on par with LingBot-World and HY-WorldPlay, and achieving 36x higher throughput.