tencent-hunyuan/hy-world-2.0 — explained in plain English
Analysis updated 2026-05-18
Reconstruct a 3D digital twin of a space from a casual video using WorldMirror 2.0
Generate a navigable 3D scene from a single text prompt or image for use in a game engine
Explore AI generated 3D worlds interactively in first person or third person mode
| tencent-hunyuan/hy-world-2.0 | nvlabs/protomotions | nvidia-nemo/datadesigner | |
|---|---|---|---|
| Stars | 1,911 | 1,945 | 1,859 |
| Language | Python | Python | Python |
| Last pushed | — | 2026-07-04 | — |
| Maintenance | — | Active | — |
| Setup difficulty | hard | hard | moderate |
| Complexity | 5/5 | 5/5 | 3/5 |
| Audience | researcher | researcher | developer |
Figures from each repo's GitHub metadata at analysis time.
Requires GPU hardware and multiple large pretrained model components, full generation pipeline not yet fully released.
HY-World 2.0 is a research project from Tencent's Hunyuan team that builds full three dimensional worlds from text, images, or video, rather than generating flat video that disappears after playback. It accepts several kinds of input, including text prompts, single images, multiple view images, and video, and outputs persistent 3D representations such as meshes and Gaussian splats that can be opened and edited in engines like Blender, Unity, Unreal Engine, or Isaac Sim. The system offers two main capabilities. World generation takes a text prompt or a single image and produces a navigable 3D scene through four stages: generating a panorama with a companion model called HY-Pano 2.0, planning a camera trajectory with a component called WorldNav, expanding the scene outward with WorldStereo 2.0, and composing the final result with WorldMirror 2.0 combined with 3D Gaussian splatting. World reconstruction takes multiple images or a casual video and, through WorldMirror 2.0, predicts depth, surface normals, camera parameters, point clouds, and Gaussian splat attributes in a single pass, effectively turning a video into a digital twin of the space it shows. The project frames its advantage over pixel based world models like Genie 3 or Cosmos around editability and permanence: the output is a real, reusable 3D asset rather than a video clip with a fixed length, it plays in real time on consumer GPUs once generated, and it supports precise physics based collision instead of imprecise character control. HY-World 2.0 also supports first person and third person interactive exploration of the worlds it produces. As of the most recent update, the project has open sourced the HY-Pano 2.0 panorama model and the WorldMirror 2.0 reconstruction model with their inference code and weights, while full world generation code and a couple of related components are still marked as coming soon. The project provides a technical report, pretrained models on Hugging Face, and demo videos showing generated scenes and reconstructions.
A research model that builds editable, persistent 3D worlds and digital twins from text, images, or video.
Mainly Python. The stack also includes Python, PyTorch, 3D Gaussian Splatting.
Setup difficulty is rated hard, with roughly 1day+ to a first successful run.
Mainly researcher.
This repo across BitVibe Labs
Verify against the repo before relying on details.