Reproduce “WorldDreamer: Towards General World Models for Video Generation via Predicting Masked Tokens” 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
Xiaofeng Wang, Zheng Zhu, Guan Huang, Boyuan Wang, Xinze Chen, Jiwen Lu
WorldDreamer introduces a world model that redefines general video generation as an unsupervised visual sequence modeling task by predicting masked visual tokens using a Transformer-based architecture. This model achieves versatile video creation and editing, including image-to-video, text-to-video, and action-to-video, with significantly faster inference speeds than existing methods.
Key insights
- Framing world modeling as an unsupervised visual sequence modeling task, similar to large language models' masked token prediction, is an effective paradigm for learning general visual dynamics.
- A Transformer architecture combining localized spatial-temporal patchwise self-attention with spatial-wise cross-attention can efficiently learn complex spatio-temporal dynamics and integrate multimodal conditioning.
- A dynamic masking strategy with consistent masks across frames, coupled with parallel sampling, enables substantial acceleration of video generation, moving beyond sequential token prediction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.