Introducing SANA-WM: A Breakthrough in Minute-Scale World Modeling with Hybrid Linear Diffusion Transformer

SANA-WM, a 2.6B-parameter open-source world model, has been developed to synthesize high-fidelity, 720p minute-scale videos with precise camera control. This model achieves visual quality on par with established industrial benchmarks while improving efficiency. Key architectural innovations include Hybrid Linear Attention for memory-efficient long-context modeling, Dual-Branch Camera Control for accurate trajectory adherence, a Two-Stage Generation Pipeline for enhanced output quality, and a Robust Annotation Pipeline for extracting high-quality action labels. SANA-WM utilizes approximately 213K public video clips for training, completes training in 15 days on 64 H100s, and efficiently generates video clips on a single GPU. In benchmark tests, it outperforms previous open-source models in action-following accuracy while achieving a significant increase in throughput.