HunyuanWorld 1.0: Tencent's Open-Source Text-to-3D World Framework

What HunyuanWorld 1.0 Generates and Who Needs It

HunyuanWorld 1.0 addresses a specific gap in the 3D content creation pipeline: generating a complete, navigable scene from either a text description or a reference image, rather than a single isolated mesh. The target audience is engineers and researchers who need scene-scale 3D geometry for virtual reality experiences, physics simulation environments, and game development prototypes.

Earlier approaches split into two unsatisfying options. Video diffusion methods could produce photorealistic scene traversals but had no underlying 3D geometry, making them impossible to re-render from new viewpoints or export to a game engine. Traditional 3D generation methods provided solid geometry but struggled with the breadth of scene types and often produced memory-intensive representations unsuited to large environments. HunyuanWorld 1.0 uses a panoramic intermediate representation to bypass both constraints.

The technical report is available at arxiv.org/abs/2507.21809, and model weights are hosted on HuggingFace under tencent/HunyuanWorld-1. The project is not distributed as a pip-installable package; setup requires the Docker environment in the docker/ directory.

The Three-Stage Pipeline: Panoramic Proxy, Semantic Layering, and Mesh Reconstruction

The generation pipeline runs in three stages, each of which corresponds to a component in the hy3dworld/ directory.

First, the framework generates a 360-degree equirectangular panorama from the input. This panorama acts as a world proxy: it encodes the full visual context of the environment in a format that diffusion models handle well. When text is the input, the system generates this panorama from the description. When an image is the input, the system extends it into a full 360-degree view.

Second, the framework applies semantic layering to that panorama. It identifies distinct regions such as sky, ground plane, structures, and individual objects, and assigns each pixel a semantic class. This decomposition is what makes the third stage possible.

Third, the layered panorama is converted to a set of 3D meshes through hierarchical reconstruction. The result is a collection of separate mesh objects rather than a single fused point cloud. The README abstract describes this design as yielding 360-degree immersive coverage, mesh formats compatible with existing graphics pipelines, and what it calls disentangled object representations for augmented interactivity. Because each object is a separate mesh, downstream applications can select, move, or replace individual scene elements without affecting the rest of the scene.

The README includes benchmark comparisons against Diffusion360, MVDiffusion, PanFusion, LayerPano3D, and WonderJourney across BRISQUE, NIQE, Q-Align, and CLIP metrics. In text-to-world generation, HunyuanWorld 1.0 achieves a BRISQUE score of 34.6 and a Q-Align score of 4.2, compared to LayerPano3D at 35.3 and 3.9 respectively.

Getting Started: Model Weights, Demo Scripts, and the Docker Environment

The repository does not include a step-by-step installation guide in the README. The recommended setup path is through the docker/ directory, which provides a containerized environment with the necessary GPU dependencies. Model weights are on HuggingFace at tencent/HunyuanWorld-1 and must be downloaded separately before running either demo.

Two demo scripts are included at the repository root. demo_panogen.py runs the panorama generation stage independently, which is useful for verifying that the model weights are loaded correctly and that the GPU is functioning as expected. demo_scenegen.py runs the full pipeline from a text prompt or image through to a 3D scene output.

Nine example input cases are provided under examples/, labelled case1/ through case9/, giving a sense of the range of input types the model handles. A browser-based model viewer ships as modelviewer.html at the root, which can load and inspect the generated scene meshes locally without additional software.

The repository is hosted at github.com/Tencent-Hunyuan/HunyuanWorld-1.0, and a web demo is available at 3d.hunyuan.tencent.com/sceneTo3D for testing the model without a local setup.

HunyuanWorld-1.0-lite: The Quantized Variant for Consumer GPUs

The base HunyuanWorld 1.0 model requires datacenter-class or high-end workstation hardware. On August 15, 2025, Tencent released HunyuanWorld-1.0-lite, a quantized version of the model. The README states that this variant supports running on consumer-grade GPUs such as the RTX 4090, making local experimentation accessible without a multi-GPU server.

The README does not include a side-by-side quality comparison between the base and lite variants, so the extent of the quality trade-off is not documented in the README. For applications where output quality is critical, testing both variants against the target use case is the only way to determine whether the lite variant meets requirements.

Structural Limits of the Panoramic Proxy Approach

The panoramic proxy approach has a topological constraint that shapes what the model can and cannot generate well. It represents a scene as a 360-degree sphere from a fixed center point. Environments that require moving through many rooms, passing through doorways, or navigating a space with a non-spherical topology do not map cleanly onto this representation. The approach works for open outdoor scenes and single-room interiors but is less suited to multi-room architectural walkthroughs, underground spaces where the sky is absent, or procedurally large game worlds.

The absence of detailed installation documentation is also a practical obstacle. The README points to Docker but does not walk through the setup steps, so first-time users must work backward from the Dockerfile to understand the dependency chain.

The repository has no versioned releases on GitHub. Users pulling from the main branch get whatever state the code is in at the time of cloning. The README news section notes that HY-World-2.0 was released on April 16, 2026, one day after the last recorded push to this repository, which indicates that active development has moved to the successor version.

HunyuanWorld 1.0 vs. LayerPano3D

LayerPano3D is the closest structural alternative: it also uses a layered panoramic representation to generate 3D scenes from images. The key difference is in output format. LayerPano3D does not produce the disentangled per-object meshes that HunyuanWorld 1.0 exports. For applications that need to move or replace scene elements after generation, HunyuanWorld 1.0's output format provides more flexibility.

The benchmark tables in the README show HunyuanWorld 1.0 outperforming LayerPano3D in every measured metric for both text-to-panorama and text-to-world tasks. In text-to-panorama generation, HunyuanWorld 1.0 scores 40.8 on BRISQUE and 4.4 on Q-Align, compared to LayerPano3D at 49.6 and 3.7. If quality metrics are the primary concern, the numbers favor HunyuanWorld 1.0. If the deployment environment requires a simpler dependency chain or a codebase with a longer release history, LayerPano3D remains an alternative worth evaluating.

Maintenance Status and Licensing

The last push to this repository was on April 15, 2026. The repository is not archived. However, the news section records that HY-World-2.0 was released on April 16, 2026, the day after the last push recorded here. This pattern is consistent with a codebase that has entered maintenance-only status while development effort moves to the 2.0 branch. The README also documents earlier follow-on releases: HunyuanWorld-1.1 on October 22, 2025, and HunyuanWorld-1.5 on December 18, 2025, each adding capabilities such as video input and real-time generation.

The license field in the repository metadata reads NOASSERTION, meaning the standard tooling did not identify a recognized open-source license identifier. Engineers intending to use this project in a commercial product should inspect the LICENSE and NOTICE files in the repository directly before proceeding. The NOTICE file is present at the root level.

Editorial conclusion

HunyuanWorld 1.0 is the right starting point for researchers who need open-source text-to-3D-world generation with explorable geometry and disentangled object meshes. Engineers targeting production deployments should first clarify the licensing situation by reading the LICENSE and NOTICE files, check whether the base or lite model fits their hardware, and consider whether HY-World-2.0 better matches current requirements given that development activity shifted to that branch in April 2026.

Frequently asked questions

Official sources

1. Issues
2. Project website
3. README
4. Tencent-Hunyuan/HunyuanWorld-1.0 on GitHub