Tencent Huan Yuan team officially announced the release and open-source of "Huan Yuan 3D World Model 2.0" (HY-World2.0) today. As a major upgrade of the multimodal world model, this version marks a transition from simple video generation to generating "interactive and editable 3D assets". The model supports various modal inputs such as text, images, and videos, and can generate, reconstruct, and simulate complete 3D spaces containing people, objects, and scenes. It also supports exporting in multiple formats such as Mesh, 3DGS, and point clouds, and seamlessly integrates with mainstream game engines like Unity and UE.

Technically, HY-World2.0 unifies the processes of spatial understanding, generation, and reconstruction. Its core upgrades include: the HY-Pano-2.0 model using an end-to-end implicit learning approach, which can achieve 360-degree panoramic mapping without requiring camera parameters; self-developed spatial agent technology combined with VLM and navmesh representation, which gives the model the ability to intelligently plan roaming trajectories; and the WorldStereo mechanism, which ensures high consistency in geometry and visual appearance between newly generated areas and existing scenes.


In addition, the upgraded WorldMirror2.0 architecture supports replicating real-world scenes, enabling the prediction of dense point clouds and camera parameters in one go, thus achieving high-precision digital twin construction.
Compared to mainstream models like Google's Genie3, the breakthrough of Huan Yuan 2.0 lies in the fact that the generated assets have realistic physical collision properties and support free exploration in character mode, significantly lowering the threshold for building game level prototypes and embodied intelligent simulation environments.
This release marks the evolution of Tencent Huan Yuan from "manipulating objects" to "creating worlds" in the field of 3D generation. Through a state-of-the-art spatial consistency memory mechanism, it pushes AI-driven 3D content creation into a practical phase.



