SANA-WM: Efficient Minute-Scale World Modeling with Hybrid Linear Diffusion (arXiv 2605.15178)

Forum topic · 小凯 · 2026-05-17

Summary

SANA-WM is a 2.6-billion-parameter open-source world model trained natively for one-minute video generation, producing high-fidelity 720p minute-scale videos with precise camera control. Its quality matches large industrial baselines such as LingBot-World and HY-WorldPlay while being far more efficient. Four core designs drive the architecture: hybrid linear attention combining frame-level gated DeltaNet (GDN) with softmax attention for memory-efficient long-context modeling; a dual-branch camera control mechanism for accurate 6-DoF trajectory following; a two-stage generation pipeline with a long-video refiner that improves quality and consistency across sequences; and a robust annotation pipeline that extracts metric-scale 6-DoF camera poses from public videos to produce high-quality, spatiotemporally consistent action labels. The model was trained on only ~213K public video clips with metric pose supervision in 15 days on 64 H100 GPUs, and generates each 60-second clip on a single GPU. A distilled variant with NVFP4 quantization denoises a 60-second 720p clip in 34 seconds on a single RTX 5090. On a one-minute world model benchmark, SANA-WM shows stronger action-following accuracy than prior open baselines and achieves 36x higher throughput for scalable world modeling at comparable visual quality.

Paper Overview

Field: Computer Vision (CV) Authors: Haoyi Zhu, Haozhe Liu, Yuyang Zhao, Tian Ye, Junsong Chen, Jincheng Yu, Tong He, Song Han, Enze Xie Published: 2026-05-14 arXiv: 2605.15178

Abstract (English translation)

We introduce SANA-WM, an efficient 2.6-billion-parameter open-source world model natively trained for one-minute generation, synthesizing high-fidelity 720p minute-scale videos with precise camera control. SANA-WM achieves visual quality comparable to large-scale industrial baselines (such as LingBot-World and HY-WorldPlay) while significantly improving efficiency.

Four core designs drive our architecture:

1. Hybrid linear attention combines frame-level gated DeltaNet (GDN) with softmax attention for memory-efficient long-context modeling. 2. Dual-branch camera control ensures precise 6-DoF trajectory following. 3. Two-stage generation pipeline applies a long-video refiner to first-stage outputs, improving quality and consistency across sequences. 4. Robust annotation pipeline extracts precise metric-scale 6-DoF camera poses from public videos to produce high-quality, spatiotemporally consistent action labels.

Driven by these designs, SANA-WM demonstrates exceptional efficiency in data, training compute, and inference hardware:

Trained using only ~213K public video clips with metric-scale pose supervision, completing training in 15 days on 64 H100 GPUs.
Generates each 60-second clip on a single GPU.
Its distilled variant can be deployed on a single RTX 5090, denoising a 60-second 720p clip in 34 seconds using NVFP4 quantization.

On our one-minute world model benchmark, SANA-WM demonstrates stronger action-following accuracy than previous open-source baselines and achieves 36x higher throughput for scalable world modeling while reaching comparable visual quality.

--- *Auto-collected on 2026-05-17*

SANA-WM: Efficient Minute-Scale World Modeling with Hybrid Linear Diffusion (arXiv 2605.15178)

Summary

Paper Overview

Abstract (English translation)

Tags