A Real 3D World Model

Tencent has released HY-World 2.0. The core point is that instead of mimicking video, it generates actual 3D assets such as 3D Gaussian Splats, meshes, point clouds.

The model aims to be a persistent world where a world, once created, can be continuously maintained and edited, and it explains that it can be imported directly into Unity, Unreal Engine, Blender, Isaac Sim.

It also claims to support physics, collision, first-person navigation, real-time rendering on consumer GPU. Unlike existing video-centric models such as Genie 3 and HY-World 1.5, it also claims to offer true 3D consistency without time limits or flickering.

The core component is WorldMirror 2.0, where a single feed-forward model predicts depth, normals, camera params, point clouds, 3DGS all at once.

The input modalities supported are as follows.

• text → 3D world
• image → 3D world
• multi-view → 3D world
• video → 3D world

Release links are posted on Hugging Face and GitHub.