ByteDance’s Doubao AI team has released VideoWorld on Github, a new video generative model in partnership with Beijing Jiaotong University and the University of Science and Technology of China.

With 300 million parameters, VideoWorld differs itself from traditional models by independently acquiring knowledge instead of relying solely on textual or tagged input data.

This allows the model to perform complex tasks like origami folding and bow tying with improved intuition.

To test its capabilities, researchers created two scenarios: a Go game to assess rule learning and reasoning, and robot simulations to evaluate planning and manipulation. Training with video demonstrations improved its ability to predict future frames.