Search papers, labs, and topics across Lattice.
This paper introduces SparSTAR, a training-free block-sparse attention mechanism designed for efficient video synthesis in autoregressive models. By scoring contiguous key blocks and executing them through a sparse path, SparSTAR improves computational efficiency while maintaining high fidelity in video generation. The method achieves a 1.6x speedup in end-to-end processing without sacrificing the quality of outputs compared to the dense InfinityStar model.
Achieving a 1.6x speedup in video synthesis while preserving output quality could redefine efficiency benchmarks in autoregressive models.
InfinityStar extends visual autoregressive generation to video through a sequence of image and clip pyramids. Its changing scale and cross-clip context, however, leave late-scale attention costly and make sparse patterns reused from diffusion or image VAR models unreliable. We introduce SparSTAR, a training-free block-sparse attention method tailored to this setting. At each expensive scale and attention head, SparSTAR scores contiguous key blocks from the current query and key activations, retains required conditioning context, and executes the selected blocks through a forward-only sparse path. We analyze cross-scale consistency within a clip, pattern persistence across clip boundaries, and quality degradation as reuse spans increasingly distant scales. Across these analyses, important key blocks shift, showing that recomputing block selection at each target scale is more reliable than reusing a transferred mask. On 720p text-to-video and image-to-video generation, SparSTAR preserves every token and refinement scale while providing about a 1.6x end-to-end speedup and maintaining VBench and paired-output reconstruction fidelity close to dense InfinityStar.
