SANA-Video 2.0: Hybrid Linear Attention with Attention Residuals for Efficient Video Generation

SANA-Video 2.0 introduces hybrid linear attention with attention residuals to reduce computational cost in video generation while preserving output quality. This method replaces quadratic attention with linear complexity, lowering FLOPs for long-sequence video synthesis.

For operators, this directly reduces GPU memory and inference latency, making video generation viable on mid-range accelerators like A100-class hardware rather than requiring H100 clusters. The attention residuals maintain coherence across frames, addressing a failure mode that previously forced teams to use full attention or accept quality loss. Builders should evaluate whether this enables real-time video editing or batch pipelines for longer clips, shifting deployment strategy from diffusion-step reduction to attention scaling. Teams currently using diffusion-based video models may find a cheaper alternative in hybrid attention, potentially altering hardware procurement and model serving architectures.