ByteDance Unveils New World Model Technologies

ByteDance demonstrates a breakthrough in the field of world models, capable of generating interactive 3D environments from text descriptions.

The new system allows users to navigate the generated space, control the camera, and interact with objects. The demonstration shows the creation of a room from text, a character flying on a paper airplane, and adding new items via command.

Key features of the demonstrated model include:

• Environment generation from a first-person text prompt
• Free adjustment of camera angle and character control
• Dynamic addition of objects (e.g., a spoon) with interaction capabilities
• Use of portals to connect different spaces
• Pre-visualization of cinematic shots, including stylization reminiscent of Stephen Chow films

Earlier in September, ByteDance introduced the research model ENUMA, which took third place in the WBench benchmark for navigation tasks with a score of 81.7. The accelerated autoregressive version of ENUMA uses only four generation steps, combining camera control, text commands, and visual references.

Simultaneously, the ActWorld project is being developed—a model with memory of actions and objects, aimed not only at navigation but also at object manipulation. ByteDance founder Zhang Yiming personally oversees the development of interactive world models based on Seedance technology.

Expected application areas include games, interactive video, and Pico VR headsets. It is expected that an intermediate model of the 2.x series will be released before the final Cidan 3.0 version this year, possibly in November.

Additional resources:

1. ENUMA researcher page: https://song630.github.io/yizhisong.github.io/
2. ActWorld project: https://interactwm.github.io/ActWorld/
3. Bloomberg report on world model development: https://www.bloomberg.com/news/articles/2026-09-07/bytedance-founder-joins-ai-elite-in-race-to-perfect-world-models