HunyuanWorld-1.0: Generate Interactive 3D Worlds from Text or Images

Overview

HunyuanWorld-1.0 is a Python project for generating scene-scale 3D worlds from text or image input. It addresses a gap between video-style generation, which can lack 3D consistency, and 3D-first methods, which can be constrained by data and memory demands.

Its pipeline generates a panoramic image, then uses semantic layering and hierarchical reconstruction to build a 3D scene. The resulting worlds are intended for exploration and interaction, with mesh export for use in graphics workflows. This is most relevant to people working on 3D content creation, virtual reality, simulation, or research into generative world models.

Key Features

• Generate panoramas from either text prompts or input images.
• Convert panoramas into layered, reconstructed 3D scenes.
• Support 360-degree immersive viewing and world exploration.
• Separate selected foreground objects into semantic layers for added interactivity.
• Export meshes for compatibility with graphics pipelines.
• Provide FP8 quantization and cache options for inference.
• Include a browser-based viewer for generated scene files.

Use Cases

• 3D artists and content creators: prototype explorable environments from a written scene description or a reference image.
• Game developers: create scene concepts and export geometry for further processing in an existing graphics pipeline.
• VR and immersive-media teams: explore panoramic environments as 360-degree worlds rather than flat images.
• Simulation and computer-vision researchers: examine a generation workflow that combines panoramic proxies with semantically layered 3D reconstruction.

Project Facts

• Language: Python
• License: NOASSERTION
• Stars: 2.9k
• Forks: 267
• Topics: 3d, 3d-generation, aigc, hunyuan3d, image-to-3d, scene-generation, text-to-3d, world-model, world-models
• Archived: false

Getting Started

The README tests the project with Python 3.10 and PyTorch 2.5.0+cu124. Start by cloning the repository and creating its Conda environment:

git clone https://github.com/Tencent-Hunyuan/HunyuanWorld-1.0.git
cd HunyuanWorld-1.0
conda env create -f docker/HunyuanWorld.yaml

The full setup includes additional dependencies and model access. See the README for installation, inference examples, and viewer instructions. Model checkpoints are linked from the Hugging Face model page.

Considerations

• This is an inference-oriented research project, not a turnkey scene editor. The documented workflow runs separate panorama-generation and scene-generation steps.
• Setup is substantial: the README lists Real-ESRGAN, ZIM, Draco, Hugging Face authentication, and a CUDA-enabled PyTorch environment.
• The repository notes that hardware limitations can prevent some generated scenes from loading in the viewer. Its lite quantized release is described as supporting consumer GPUs such as the 4090, but hardware suitability depends on the workload.
• The repository input reports the license as NOASSERTION, so check the repository and model terms before adopting it in a project.
• The repository points to newer Hunyuan world-model releases. For new work, compare this version with those successors to ensure its workflow and capabilities fit your needs.