DCARL: A Divide-and-Conquer Framework for Autoregressive Long-Trajectory Video Generation
TLDR
A divide-and-conquer autoregressive framework for long-trajectory video generation combining keyframe and interpolation generators.
Reasoning
Strengths include a novel combination of divide-and-conquer with autoregressive generation to address visual drift and controllability, achieving superior results on long videos. Weaknesses are that the paper focuses on video generation rather than explicitly modeling world dynamics or interactivity, limiting its direct relevance to world model keywords.
Read-first score
Read-first score 42.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 13.
Field roles
Rank sensitivity
Stability: volatile; rank range: 472.
Keyword Scores
Deep Analysis
Innovations
- Divide-and-conquer autoregressive framework combining structural stability of divide-and-conquer with high-fidelity generation of video diffusion models
- Dedicated Keyframe Generator trained without temporal compression to establish long-range globally consistent structural anchors
- Interpolation Generator that synthesizes dense frames autoregressively with overlapping segments, using keyframes for global context and a single clean preceding frame for local coherence
Methodology
DCARL employs a two-stage generation process: first, a Keyframe Generator produces long-range structural anchors without temporal compression; then, an Interpolation Generator autoregressively fills dense frames using overlapping segments, conditioned on keyframes for global context and a single preceding frame for local coherence. The model is trained on a large-scale internet long-trajectory video dataset.
Key Results
DCARL achieves superior visual quality (lower FID and FVD) and camera adherence (lower ATE and ARE) compared to state-of-the-art autoregressive and divide-and-conquer baselines, enabling stable and high-fidelity generation for long trajectory videos up to 32 seconds.