INTERACTIVE WORLD MODELS · PROJECT PAGE

H3-World: Turning Language Understanding
into World Control

H3-World enables character and camera control across diverse visual environments.

Danze Chen1,2,*Zeqing Wang1,2,*Ziyue Lin3Xingyi Yang3,*Yeying Jin1,2,*
1 Tencent2 National University of Singapore3 The Hong Kong Polytechnic University

♣ Research internship at Tencent  ·  * Corresponding authors  ·  ◆ Project Leader

H3-World across four visual environments
H3-World turns language instructions into controllable character, camera, and world behavior.
01

World Control Demos

Language becomes direct, visible control across characters, cameras, and environments.

01

Neon alley · fast pan right

First-person camera control in a neon-lit city.

02

Lantern cave · strafe left

First-person lateral movement in a complex cave environment.

03

Dune oasis · strafe right

Third-person navigation across an open desert scene.

04

Cathedral paladin · tilt up

Third-person camera motion in a detailed fantasy interior.

From understanding
to control.

A world model should do more than recognize what is happening. It should understand what a person wants to happen next.

H3-World provides one language for character, viewpoint, motion, and environment, keeping directions connected instead of treating them as isolated effects.

CORE IDEALanguage is the control surface for an interactive visual world.
03

Method

H3-World learns language-aligned actions while preserving the generative prior of the world model.

H3-World method diagram
Figure 1. H3-World packs visual tokens and latent-aligned action conditions into a unified world-control model.
04

Temporally specified control

Actions can change over time while the generated sequence remains visually coherent.

Temporal world control examples
05

Action controllability

Language-aligned action conditioning steers motion without discarding scene content.

Action conditioning diagram
CharacterIdentity, action, and purpose.
CameraViewpoint, movement, and composition.
EnvironmentPlace, light, weather, and response.
06

Compositional generalization

H3-World transfers control to unseen combinations of visual environments and actions.

Compositional generalization results
07

Citation

If H3-World supports your work, cite the project.

@article{chen2026h3world,
  title={H3-World: Turning Language Understanding into World Control},
  author={Chen, Danze and Wang, Zeqing and Lin, Ziyue and Yang, Xingyi and Jin, Yeying},
  year={2026}
}
H3-World | Turn language into world control