Chorus II: Cross-Request Sparsity Reuse for Efficient Image-to-Video Generation

Chorus II is a framework designed to accelerate the serving of diffusion-based image-to-video (I2V) models. Because I2V workloads often involve recurring subjects, layouts, or templates, the authors observe that these similar requests share highly consistent sparse attention patterns. By treating these sparse masks as cacheable artifacts, Chorus II avoids the expensive online mask-prediction overhead typically required by dynamic sparse attention methods, particularly in few-step distilled models where such overhead is difficult to amortize.

The core of the framework is "sparsity reuse," where a new request retrieves a high-quality sparse mask from a similar historical request (indexed via joint text-image embedding similarity). To ensure robustness, the authors implement a safety fallback mechanism that includes a minimum top-k constraint and a block-pair visit refresh, ensuring that even with slight distributional drift between requests, the model maintains sufficient information flow.

For more aggressive acceleration, the framework offers an optional "downsampled latent feature reuse" mode. Unlike previous methods that simply skip redundant regions—which can lead to boundary artifacts and tearing—Chorus II uses spatiotemporal downsampling to compute a lightweight context for these regions. This preserves global coherence while significantly reducing computational costs. Finally, a "guidance enhancement" strategy is applied to mitigate semantic drift and identity inconsistency that can arise from reusing features or masks across non-identical requests.

Experiments on few-step distilled models (such as Wan 2.2) demonstrate that the default sparsity reuse configuration achieves a 2.16x speedup in FP8. When the optional feature reuse extension is enabled, this speedup increases to 2.59x. By shifting the burden of mask construction from online computation to a retrieval-based cache, Chorus II provides a scalable solution for high-fidelity video generation services.