SANA-Video 2.0 (nvlabs.github.io)

NVIDIA Research has unveiled SANA-Video 2.0, a revolutionary hybrid video diffusion transformer engineered for efficient video generation at scales of 5B and 14B parameters. This innovative architecture enables the generation of high-quality videos up to 720p resolution using a single GPU, achieving comparable quality to full-softmax video diffusion transformers (DiTs) while benefitting from the scalability of linear attention mechanisms. By incorporating Hybrid Linear-Softmax Attention, which combines gated linear attention with periodic gated-softmax anchors, the model effectively overcomes the limitations of quadratic attention and enhances long-sequence processing efficiencies. Significantly, SANA-Video 2.0 demonstrates remarkable performance with a VBench score of 84.30 at 480p resolution in just 13.2 seconds, and its compiled forward pass is 3.2 times quicker than a full-softmax counterpart at 720p/60s. Advanced optimizations such as full-stack Sol-Engine enhancements further improve performance, making the 5B pipeline 120 times faster than prior models. This breakthrough not only recovers the expressiveness of softmax but does so at a substantially reduced computational cost, paving the way for scalable, high-resolution video creation that could transform the landscape of AI-generated media.