I am a co-first author of CineFlow, advised by Prof. Fan Lai at UIUC.
CineFlow reframes long-form video diffusion as a dependency-driven execution problem. Instead of treating a prompt as one monolithic generation job, it extracts sparse semantic dependencies among scenes, entities, and events, then executes independent work in parallel while retaining the coordination needed for visual consistency.
System design
- Semantic-Dependency Compiler — compiles free-form prompts into weighted DAGs of schedulable generation units.
- Trajectory-Aware Fuser — streams latent features across dependent tasks and prunes interactions as trajectories diverge.
- Dependency-Aware Scheduler — combines critical-path scheduling, backfilling, and adaptive synchronization.
Built atop xDiT and evaluated across Wan2.2-5B, CogVideoX-5B, and HunyuanVideo on 8×H100s, CineFlow achieves 1.7–5.5× end-to-end speedup, 5.4–17.3% higher VBench overall scores, and 1.30–2.02× lower P90 latency in scheduler comparisons.