Synthesizing consistent and coherent long video remains a fundamental challenge. Existing methods suffer from semantic drift and narrative collapse over long horizons. We present A²RD, an Agentic Auto-Regressive Diffusion architecture that enables video diffusion models to synthesize and self-improve long videos autoregressively, enforcing temporal consistency and narrative coherence. A²RD is training-free and built upon three pillars:
- Multimodal Video Memory: Disentangles each segment into Textual States, Frames, and Videos to maintain both visual and narrative coherence.
- Adaptive Segment Generation: Adaptively selects between extrapolation and interpolation per segment based on narrative structure.
- Hierarchical Test-Time Self-Improvement (HITS): Self-improves boundary frames and full segments hierarchically to prevent error cascading.
Across public and LVBench-C benchmarks spanning one- to ten-minute videos, A²RD outperforms state-of-the-art baselines by up to 30% in consistency and 20% in narrative coherence.
For full video gallery, please visit our Project Page.
Examples of how A²RD preserves narrative coherence and temporal consistency:
🔍 Temporal Consistency
🔍 Narrative Progression
We introduce LVBench-C (Long Video Bench-Challenge), a benchmark designed to stress-test temporal consistency under complex scenarios where entities and environments appear, disappear, and reappear across long horizons with optional state changes.
Coming soon.
This project page is based on the Nerfies template.