Accelerating Text-to-Video Generation with Calibrated Sparse Attention

Introduces CalibAtt, a training-free method for accelerating text-to-video generation using calibrated sparse attention. CalibAtt identifies block-level sparsity and repetition patterns in transformer-based diffusion models that are stable across inputs, compiling them into optimized attention operations. This allows for skipping unselected connections at inference time, achieving up to 1.58x speedup while maintaining quality and text-video alignment.