Reproduce “DCARL: A Divide-and-Conquer Framework for Autoregressive Long-Trajectory Video Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Junyi Ouyang, Wenbin Teng, Gonglin Chen, Yajie Zhao, Haiwei Chen
DCARL introduces a two-stage, divide-and-conquer framework for autoregressive long-trajectory video generation, addressing visual drift and enhancing precise camera control. The method achieved an FID of 19.2 and FVD of 203.7 on 32-second videos from the ODV-YT dataset, alongside improved camera adherence metrics (ATE 0.237, ARE 7.669) compared to prior baselines.
Key insights
- Employing sparse, globally consistent keyframes as anchors effectively bounds error accumulation in autoregressive generation, significantly improving long-term consistency and controllability.
- Perturbing keyframe latents with consistent noise during interpolation forces the model to learn genuine motion dynamics and trajectory adherence, preventing 'copy-paste' shortcuts and unnatural transitions.
- Decoupling the generation into global structure and local detail stages allows for high spatial fidelity in sparse anchors and smooth temporal continuity in dense segments, enhancing overall video quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.