SANA-WM: Efficient Minute-Scale World Modeling with Hybrid Linear Diffusion Transformer

2026-05-14 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision & Pattern Recognition, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

SANA-WM is an efficient 2.6B-parameter open-source world model designed for generating high-fidelity, 720p, minute-scale videos with precise camera control. It achieves visual quality comparable to industrial baselines like LingBot-World and HY-WorldPlay, while significantly enhancing efficiency. Key architectural innovations include Hybrid Linear Attention for memory-efficient long-context modeling, Dual-Branch Camera Control for 6-DoF trajectory adherence, and a Two-Stage Generation Pipeline with a long-video refiner. SANA-WM also utilizes a Robust Annotation Pipeline to extract accurate metric-scale 6-DoF camera poses from public videos. The model was trained in 15 days on 64 H100s using approximately $213K public video clips and can generate a 60s clip on a single GPU, with a distilled variant running on an RTX 5090 using NVFP4 quantization.

Key takeaway

For research scientists developing world models, SANA-WM demonstrates that combining hybrid linear attention with a two-stage generation pipeline can significantly improve efficiency and throughput without sacrificing visual quality. You should consider these architectural designs to reduce training compute and enable single-GPU inference for minute-scale video generation, especially when precise camera control is critical.

Key insights

SANA-WM offers efficient, high-fidelity minute-scale video generation with precise camera control via hybrid attention and a two-stage pipeline.

Principles

• Hybrid attention improves long-context modeling efficiency.
• Dual-branch control ensures precise camera trajectory.
• Two-stage refinement enhances video quality and consistency.

Method

SANA-WM employs Hybrid Linear Attention, Dual-Branch Camera Control, and a Two-Stage Generation Pipeline, supported by a Robust Annotation Pipeline for 6-DoF camera pose extraction from public videos.

In practice

• Generate 60s 720p videos on a single GPU.
• Deploy distilled variant on RTX 5090 with NVFP4.
• Achieve 36x higher throughput for world modeling.

Topics

• SANA-WM
• World Modeling
• Hybrid Linear Diffusion Transformer
• Minute-Scale Video Generation
• Camera Control

Best for: Research Scientist, AI Scientist, Computer Vision Engineer, Machine Learning Engineer

Related on AIssential

• NVIDIA Introduces SANA-WM: A 2.6B-Parameter Open-Source World Model That Generates Minute-Scale 720p Video on a Single GPU — SANA-WM is an efficient 2.6B-parameter open-source world model generating 720p video on a single GPU.
• Training-free sparse attention based on cumulative energy filtering — Dynamic thresholding for sparse attention simultaneously optimizes sparsity and accuracy in Diffusion Transformers for video generation.
• A Scale-Adaptive Framework for Joint Spatiotemporal Super-Resolution with Diffusion Models — A scale-adaptive framework reuses one architecture for spatiotemporal super-resolution across diverse scaling factors.
• ImageWAM: Do World Action Models Really Need Video Generation, or Just Image Editing? — Image editing offers a more efficient and focused alternative to video generation for robot action prediction in World Action Models.
• GenLCA: 3D Diffusion for Full-Body Avatars from In-the-Wild Videos — GenLCA generates photorealistic 3D avatars from partial 2D video data using a novel visibility-aware diffusion training strategy.
• Diffusion Transformer World-Action Model for AV Scene Prediction — Diffusion Transformers predict realistic AV future scenes, outperforming regression on perception metrics despite distortion metric bias.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.