HunyuanWorld 1.0: the last commit landed the day before the successor was announced
Generating Immersive, Explorable, and Interactive 3D Worlds from Words or Pixels with Hunyuan3D World Model
At a glance
- What is it?
- HunyuanWorld 1.0 generates a semantically layered 3D mesh from text or images, and its four benchmark tables report no geometric metric at all despite the text claiming geometric consistency. The repository itself stopped moving in April 2026, one day before a newer model was announced in a different repository.
- Who is it for?
- HunyuanWorld 1.0 is worth reading if you want the mesh export path and the 360 degree panoramic proxy idea in code you can inspect, and it is a poor choice for a new project, because the team has moved through six model generations in nine months and this repository stopped receiving commits the day after HY-World-2.0 was announced elsewhere.
- Can I use it commercially?
- Check first. The repository uses a licence we do not classify automatically, so read its LICENSE file before any commercial use.
- Is it still maintained?
- Activity is slowing. The repository last received commits 7 months ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- Tencent-Hunyuan/HunyuanWorld-1.02,969 ★ · +30
| Repository | 2026-10-01 | 2026-11-15 |
|---|---|---|
| Tencent-Hunyuan/HunyuanWorld-1.0 | 2.9k | 3k |
30 Stars in the last 45 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
The last push was 2026-04-15 and 2.0 was announced on April 16
The news block at the top of the readme is a version history for a family rather than for one repository, and reading it in order makes this checkout's status obvious. HunyuanWorld 1.0 was released on July 26, 2025, alongside a technical report. A quantized version called HunyuanWorld-1.0-lite followed on August 15, 2025, with the note that it now supports running on consumer-grade GPUs such as a 4090. An RGB-D video diffusion model called HunyuanWorld-Voyager arrived on September 2, 2025. HunyuanWorld-1.1, named WorldMirror, came on October 22, 2025 and works from videos or multi-view images. HunyuanWorld-1.5, named WorldPlay, followed on December 18, 2025 for real-time world creation. And on April 16, 2026 came HY-World-2.0. The last push recorded on this repository's default branch is dated 2026-04-15. The repository was archived in neither sense, but the sequence says plainly that this is the first generation and that development moved elsewhere.
The licence field reads NOASSERTION while the tree carries LICENSE and NOTICE
The repository's licence is the one thing about this project that cannot be answered from the page, and the reason is a genuine inconsistency rather than an omission. The licence metadata for the repository reads NOASSERTION, which means the automatic detector did not recognise a licence it could name. The top level of the tree, however, contains both a LICENSE file and a separate NOTICE file, alongside the readme, the Chinese readme, the two demo scripts, the hy3dworld/ package, a docker/ directory, a scripts/ directory, assets/, examples/ and a modelviewer.html. Neither file is linked from the readme, no licence name is written in the text, and nothing on the page says which of the two files governs the weights, the code, or both. So a user who needs to know whether the output can be used commercially has to open the repository and read two documents, and this page cannot be used to answer it.
The performance paragraph claims geometric consistency and no metric measures it
The sentence introducing the results says the numerical results indicate that HunyuanWorld 1.0 surpasses baselines in visual quality and geometric consistency. The four tables that follow measure neither of those two things directly. Their columns are BRISQUE and NIQE, which are no-reference image quality scores where a lower number is better, Q-Align, which scores how well an image matches its prompt where higher is better, and either CLIP-T or CLIP-I, a CLIP similarity score against the text or the input image, again higher is better. Every one of those is an image-space measure. Geometric consistency, the property the abstract says 3D based methods provide and video based methods lack, has no column in any of the four tables. The claim and the evidence are about different properties, and the readme does not separate them.
The winning margin falls from about seven BRISQUE to under one in the world tables
The size of the reported improvement depends heavily on which table you read, and the tables are not comparable to each other. In text to panorama generation, HunyuanWorld 1.0 scores 40.8 on BRISQUE against 47.9 for the best baseline MVDiffusion and 49.6 for LayerPano3D, a lead of roughly seven points. In text to world generation, the same column reads 34.6 against 35.3 for LayerPano3D, a lead of seven tenths of a point. NIQE in that table is 4.3 against 4.8. Q-Align is 4.2 against 3.9. So the panorama stage shows a clear margin and the final world stage shows a narrow one, on the same metrics and against overlapping baselines. The image to panorama table sits in between at 45.2 against 47.7. Whether a seventh of a point on BRISQUE is a meaningful gap is not something the page takes a position on.
Each table selects a different set of baselines
Four tables, four different comparison sets, and the page gives no reason for the variation. Text to panorama generation compares against Diffusion360, MVDiffusion, PanFusion and LayerPano3D, four baselines. Image to panorama generation compares against Diffusion360 and MVDiffusion only, so PanFusion and LayerPano3D are absent from a table measuring the same kind of output. Text to world generation compares against Director3D and LayerPano3D, so the panorama methods are gone and one new one appears. Image to world generation compares against WonderJourney and DimensionX. Read as a set, each table is a fair comparison against the methods its authors considered relevant to that stage, but it means the headline about surpassing baselines rests on a different field of competitors each time. No third party reproduction is mentioned, and the readme attributes the evaluation to the authors in the first person.
The community table has four column headers and no rows
Small artifacts are worth naming because they are what a reader sees first. The link row under the news block contains three commented-out markdown reference definitions, each of them a disabled link pointing at nothing. Below the community invitation there is a table with four headers, Wechat Group, Xiaohongshu, X and Discord, and a single body row that is empty in all four cells. So the page invites readers to join a Wechat group and a Discord server, links Discord and X in the header row above, and leaves the table itself blank. The rest of the outbound links are complete: the hosted demo at the scene to 3D page, the model weights on Hugging Face under tencent/HunyuanWorld-1, a model gallery, the arxiv report, and the Chinese readme, which the page offers as 中文阅读 at the top.
Two demo scripts at the root and a single file model viewer
The repository's shape follows the pipeline described in the architecture paragraph, which names three stages: panoramic proxy generation, semantic layering, and hierarchical 3D reconstruction. Two runnable scripts sit at the top level rather than under scripts/. demo_panogen.py corresponds to the first stage and demo_scenegen.py to the last, so the two halves of the pipeline are runnable separately and that separation is visible from the file names alone. The package itself is hy3dworld/. Then there is modelviewer.html, a single file in the root that is presumably where the exported mesh is looked at, which is the only viewer the tree provides. Examples/ holds nine directories, case1 through case9. There is also a docker/ directory whose contents are not described anywhere on the page, and no command, no dependency list and no hardware requirement appears in the visible portion of the readme, which is where a new user has to look first.
Editorial conclusion
HunyuanWorld 1.0 is worth reading if you want the mesh export path and the 360 degree panoramic proxy idea in code you can inspect, and it is a poor choice for a new project, because the team has moved through six model generations in nine months and this repository stopped receiving commits the day after HY-World-2.0 was announced elsewhere. Before using anything from it, settle the licensing question yourself: the repository's licence field is unreadable and the tree carries both a LICENSE and a NOTICE file with nothing on the page saying which terms govern the weights. And read the benchmarks as what they are, four tables of no-reference image quality and prompt alignment scores produced by the authors, with no metric that measures the geometric consistency the surrounding paragraph claims.
Frequently asked questions
hunyuan world 1.0 how to use
The repository ships two runnable scripts at its root, demo_panogen.py for panorama generation and demo_scenegen.py for scene generation, plus a docker/ directory and a modelviewer.html for looking at exported meshes. The usage section of the readme with the commands and dependencies is not part of the visible page.
What licence applies to HunyuanWorld 1.0?
This page does not say. The repository's licence field reads NOASSERTION, while the top level of the tree contains both a LICENSE file and a separate NOTICE file, neither of which is linked from the readme, and no licence name is written in the text.
What benchmarks does HunyuanWorld 1.0 report?
Four tables: text to panorama generation, image to panorama generation, text to world generation and image to world generation. The metrics are BRISQUE and NIQE for no-reference image quality, where lower is better, plus Q-Align for prompt alignment and CLIP-T or CLIP-I for CLIP similarity, where higher is better.
How does HunyuanWorld 1.0 relate to HunyuanWorld 2.0?
They are separate generations in separate repositories. The news list records HY-World-2.0 on April 16, 2026, after HunyuanWorld-1.5 WorldPlay in December 2025, HunyuanWorld-1.1 WorldMirror in October 2025, HunyuanWorld-Voyager in September 2025, and a HunyuanWorld-1.0-lite quantization in August 2025. The last push on this repository is dated 2026-04-15.
What does HunyuanWorld 1.0 actually produce?
A semantically layered 3D mesh representation that uses panoramic images as 360 degree world proxies for semantic aware world decomposition and reconstruction, from text or image conditions. The three stated advantages are 360 degree immersive experiences via panoramic world proxies, mesh export for existing graphics pipelines, and disentangled object representations for added interactivity.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/tencent-hunyuan-hunyuanworld-1-0)Star history chart for your README
[](https://hysenlabs.com/projects/tencent-hunyuan-hunyuanworld-1-0)Was this useful?