Reproduce “MagicWorld: Towards Long-Horizon Stability for Interactive Video World Exploration” 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
Guangyuan Li, Bo Li, Jinwei Chen, Xiaobin Hu, Lei Zhao, Peng-Tao Jiang
MagicWorld presents an interactive video world model, based on a Video Diffusion Transformer, that improves long-horizon stability and realistic motion for dynamic subjects during scene exploration. It achieved an overall VBench score of 0.8547, outperforming state-of-the-art methods while maintaining competitive inference speed, and introduced the RealWM120K dataset to support real-world dynamic scene modeling.
Key insights
- Explicit motion supervision via optical flow and geometric priors from 3D scene reconstruction are crucial for preserving dynamic subject behavior and structural integrity across interactions.
- Aggregating distribution matching distillation losses over multiple interactive steps and incorporating visual and motion quality rewards during training is more effective for mitigating long-horizon error accumulation than per-step optimization.
- Retrieving and injecting semantically similar latent states from a history cache significantly enhances viewpoint consistency and reduces drift in long-term interactive generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.