Tencent open-sources HY-World 2.0 to turn prompts into editable 3D scenes

HY-World 2.0 is an open-source multimodal 3D world model developed by Tencent's Hunyuan team. Basically, instead of generating pixels you can only watch like a movie, it turns text, images, or casual videos into real, editable 3D assets that drop straight into game engines.

Whether you feed it a single text prompt, a reference photo, multi-view captures, or a video clip, it outputs standard 3D representations like polygon meshes and 3D Gaussian Splatting (3DGS) that persist permanently.

What it does

Most existing world models—like Genie 3, Cosmos, or even earlier experiments like HY-World 1.5—generate pixel streams frame by frame. That approach usually means heavy inference lag, cross-view flickering, and interaction limits capped at about a minute.

HY-World 2.0 flips the paradigm by treating world modeling as game level building rather than video streaming.

Once generated, the world is yours to keep. You can immediately import the assets into Blender, Unity, Unreal Engine, or NVIDIA's Isaac Sim for further tuning. Because they are true 3D assets, rendering on consumer GPUs takes almost zero extra AI compute.

It also lets you step inside immediately. The system supports first-person navigation and third-person character control across AI-generated streets and interiors, complete with physics-based collision detection so you do not phase through walls.

Beyond generation, it handles real-world reconstruction through an integrated model called WorldMirror 2.0 (~1.2B parameters). Feed it multi-view photos or a casual smartphone video, and in a single forward pass it predicts depth, surface normals, camera poses, dense point clouds, and 3DGS attributes across 50K to 500K pixel resolutions.

Under the hood and open release

HY-World 2.0 builds full worlds using a four-stage pipeline: panorama generation with HY-Pano 2.0, camera trajectory planning with WorldNav, scene expansion via WorldStereo 2.0, and final composition combining WorldMirror 2.0 with Gaussian splat learning.

Tencent has made the project open source on GitHub under Tencent-Hunyuan/HY-World-2.0, sharing technical reports, inference scripts, and Hugging Face checkpoints. The WorldMirror 2.0 weights, HY-Pano 2.0, and WorldStereo 2.0 modules are being rolled out alongside interactive demos.

Why it matters

If you build games, design virtual sets, or train embodied AI robots, this is one of the most practical world models to land this year. Instead of being stuck with fleeting video frames, you get pipeline-ready meshes and point clouds you can actually edit, light, and simulate.

We think Tencent made a remarkably clever call here. While most labs are chasing cinematic video generation, the 3D industry desperately needed usable geometric primitives. HY-World 2.0 connects generative AI directly to standard game production workflows, showing that the real future of world models lies in playable assets, not just pre-rendered clips.