HY-World 2.0: Generate and Reconstruct Persistent 3D Worlds

HY-World 2.0 is a multimodal system for world generation, reconstruction, and simulation. A text prompt or single image can seed an entirely new navigable environment; a set of photos or a casual video can become a reconstructed digital scene.

The important difference is the output. HY-World 2.0 produces spatial assets—3D Gaussian Splats, meshes, point clouds, depth, normals, and camera estimates—so the result can be revisited from new viewpoints instead of existing only as a fixed camera sequence.

World generation begins with HY-Pano 2.0, which creates a 360-degree panorama from text or an image. WorldNav parses the scene and plans obstacle-aware camera paths; WorldStereo 2.0 generates consistent keyframes along those paths while maintaining scene memory.

WorldMirror 2.0 then estimates the geometry and camera information needed for composition, and 3DGS optimization exports the final Gaussian Splatting world. The public command-line workflow breaks this into trajectory planning, trajectory rendering, world expansion, GS data preparation, and 3DGS training.

For reconstruction, WorldMirror 2.0 accepts multi-view images or video and predicts dense point clouds, depth maps, surface normals, camera intrinsics and extrinsics, plus 3DGS attributes in one feed-forward pass. Optional camera and depth priors can be injected when they are already available.

The repository includes single-GPU and multi-GPU inference, a Python pipeline, a CLI, and a Gradio app that previews Gaussian Splats, point clouds, depth, normals, and cameras in the browser. Flexible-resolution inference is documented from roughly 50K to 500K pixels.

A video world model predicts frames while the camera moves. HY-World 2.0 builds a persistent scene representation, so the world remains available after generation, stays spatially consistent across views, and can be rendered repeatedly without rerunning a generative model for every frame.

That makes the output useful beyond a demo clip: artists can edit it, developers can combine it with game logic, and simulation teams can use the same spatial environment for repeated navigation and testing.

The official examples show both first-person navigation and third-person character exploration. Physics-based collision gives characters and cameras a usable relationship with streets, rooms, buildings, and terrain instead of simply playing a prerecorded fly-through.

HY-World 2.0 assets are designed for downstream 3D workflows and can be brought into tools such as Blender, Unity, Unreal Engine, and Isaac Sim. Meshes fit conventional editing and collision pipelines; Gaussian Splats preserve rich appearance for high-fidelity novel-view rendering.

Tencent has released the full generation inference path, WorldMirror 2.0 reconstruction, HY-Pano 2.0, WorldStereo 2.0, model weights, and technical documentation. The model zoo lists WorldMirror 2.0 at about 1.2B parameters, HY-Pano 2.0 at about 80B, a smaller HY-Pano 2-Qwen at about 425M, and WorldStereo 2.0 at about 17B.

The recommended environment is CUDA 12.8 with Python 3.11 or newer. Reconstruction has the simpler starting path; full world generation adds large models, navigation components, multi-stage rendering, and 3DGS training. The official hosted product received a HY World 2.1 update in July 2026, while the public report and repository document the HY-World 2.0 open-source stack.