Chorus II speeds up image-to-video generation 2.16x via sparsity reuse
Reusing attention masks across similar requests cuts compute with no quality loss.
Chorus II tackles the computational bottleneck of serving diffusion models for image-to-video (I2V) generation at scale. The key insight: real-world I2V workloads often contain repetitive requests—same effect templates, subjects, or shot layouts—that share consistent sparse attention patterns. Instead of predicting attention masks from scratch for each request, Chorus II reuses high-quality sparse masks from historical similar requests, slashing online overhead. On default settings, this sparsity reuse alone delivers a 2.16× speedup without degrading output quality.
Beyond sparsity reuse, the framework optionally extends to feature reuse—applying downsampled computation on redundant spatiotemporal regions while mitigating boundary artifacts. A lightweight guidance enhancement step reinforces image/text conditioning after reuse, preventing semantic drift. The method is evaluated on standard I2V benchmarks, showing no loss in generation fidelity. This makes high-throughput, low-latency deployment of I2V models significantly more practical for applications like video editing, social media effects, and automated content creation.
- Chorus II reuses sparse attention masks from similar historical requests, avoiding per-request mask prediction overhead.
- Achieves 2.16× speedup with no compromise on generation quality in its default configuration.
- Optional feature reuse and guidance enhancement further improve efficiency while maintaining condition adherence.
Why It Matters
Makes image-to-video AI practical at scale by drastically cutting compute cost for repeated content patterns.