Long-Form Video Generation

A²RD is a training-free agentic auto-regressive diffusion architecture that separates generation from consistency correction to reduce semantic drift and narrative collapse in long-form videos. At its core is a Retrieve–Synthesize–Refine–Update loop that generates segment by segment and self-improves.

The core design has three components.

• Multimodal Video Memory stores text states, boundary frames, and full segments together to track characters, environments, and camera flow.
• Adaptive Segment Generation chooses extrapolation when natural progression is needed, and interpolation when consistency between start and end matters.
• Hierarchical Test-Time Self-Improvement (HITS) corrects stepwise from boundary frames to full segments to reduce error propagation.

Operation begins with memory initialization, which analyzes the narrative to identify entities and environments, builds a dependency graph, and creates a global reference frame. Then for each segment, it retrieves context from memory, selects a generation mode, synthesizes boundary frames and video, refines with HITS, and then updates memory.

The authors also propose LVBench-C, a benchmark of multi-scene stories at 3-minute, 5-minute, and 10-minute scales, featuring nonlinear transitions where characters and environments appear, disappear, and reappear. The authors report that on public benchmarks and LVBench-C's 1-to-10-minute videos, A²RD improved consistency by up to 30% and narrative coherence by up to 20% over existing methods.