Hunyuan 3D
Last reviewed
May 16, 2026
Sources
30 citations
Review status
Source-backed
Revision
v1 ยท 3,085 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 16, 2026
Sources
30 citations
Review status
Source-backed
Revision
v1 ยท 3,085 words
Add missing citations, update stale details, or suggest a clearer explanation.
Hunyuan 3D is a family of generative artificial intelligence models and an associated creation engine built by Tencent for producing three dimensional assets from text prompts, single images, sketches, and other conditioning inputs. It belongs to the broader Hunyuan suite of foundation models, which also includes large language models, the HunyuanVideo text to video system, and the HunyuanImage diffusion model. The lineage began in November 2024 with the open release of Hunyuan3D 1.0 and progressed through Hunyuan3D 2.0, 2.1, 2.5, and 3.0 over the course of 2025. On 25 November 2025 Tencent announced the global launch of the Hunyuan 3D creation engine and an English language API on Tencent Cloud, opening the system to creators outside Mainland China for the first time.
At the time of launch the project had crossed three million cumulative downloads on Hugging Face across its open weight releases, making it one of the most widely adopted open 3D foundation models. Tencent claims that more than 150 enterprises in Mainland China had integrated the technology before the global rollout, with Unity China, the 3D printer manufacturer Bambu Lab, and the AI content platform Liblib among the named partners. The newest production model, Hunyuan 3D 3.0, supports geometric resolutions of 1,536 cubed voxels and produces meshes intended for direct use in game engines, additive manufacturing, virtual reality, and e commerce visualization.
Tencent began publishing generative AI research under the Hunyuan brand in 2023. By late 2024 the company had shipped a billion parameter class text to image model, a Mixture of Experts language model, and the HunyuanVideo system for short clip generation. The 3D track grew out of internal needs at Tencent Games, where the company estimated that hand modelling a single game ready asset took an experienced artist between four and eight hours, and where the production of large open world scenes required tens of thousands of such assets. The Hunyuan3D project set out to compress that cycle to seconds while still emitting meshes and textures that fit existing digital content creation pipelines.
The project entered a crowded field. Stability AI had released Stable Zero123 and TripoSR earlier in 2024, the startup Tripo AI had shipped commercial text to 3D systems, Meshy had built a subscription product around image to 3D, and the academic Trellis system from Microsoft Research had demonstrated structured latent diffusion for shape and texture. Tencent's contribution was scale: large diffusion transformers, a paired texture model, and an open licence that permitted commercial use across most of the world.
The table below lists the publicly disclosed Hunyuan 3D releases through May 2026. Dates refer to the first public availability of either weights, a technical report, or a hosted API.
| Version | Released | Key change |
|---|---|---|
| Hunyuan3D 1.0 | November 2024 | Two stage framework, multi view diffusion plus feed forward reconstruction, text and image input |
| Hunyuan3D 2.0 | 21 January 2025 | Hunyuan3D DiT for shape and Hunyuan3D Paint for texture; first community licence release |
| DiT v2 0 Fast | 3 February 2025 | Distilled guidance variant for faster sampling |
| Hunyuan3D 2.0 Turbo | 19 March 2025 | Step distilled checkpoints for production latency |
| Hunyuan3D 2.5 | 23 April 2025 | 10 billion parameter shape model, 1,024 cubed geometric resolution, 4K PBR textures |
| Hunyuan3D 2.1 | 13 June 2025 | First fully open source PBR pipeline, training code and full weights released |
| Hunyuan3D PolyGen | 8 July 2025 | Autoregressive mesh model with clean quad and triangle topology |
| HunyuanWorld 1.0 | July 2025 | Open source 3D world generation from text or image |
| Hunyuan3D Omni | September 2025 | ControlNet style multi condition input including pose, point cloud, voxel, and bounding box |
| Hunyuan3D 3.0 | 16 September 2025 | 1,536 cubed voxel resolution, 3.6 billion voxels, rough to fine 3D DiT |
| Hunyuan 3D Studio | September 2025 | Invitation only end to end pipeline with UV unwrapping, rigging, and skinning |
| Hunyuan 3D creation engine global launch | 25 November 2025 | Public web platform plus English Tencent Cloud API |
| HunyuanWorld 1.5 (WorldPlay) | December 2025 | Real time world generation for interactive playback |
| HunyuanWorld 2.0 | April 2026 | Updated 3D world model |
The core Hunyuan3D pipeline separates geometry from appearance. A shape model produces an untextured mesh, and a texture model paints that mesh with view consistent material maps. This decoupling is similar to the design used by Trellis, CRM, and other contemporary systems, but Tencent scaled both stages to large diffusion transformer backbones.
The shape model is named Hunyuan3D DiT. It is a flow based diffusion transformer that conditions on a reference image or a text embedding and denoises an implicit volumetric representation. The 1.0 release used a multi view diffusion stage followed by a feed forward sparse reconstruction network, but starting with 2.0 the project moved to a single stage native 3D diffusion. Across versions Tencent shipped several sizes: a 0.6 billion parameter mini variant, a 1.1 billion parameter standard variant in the 2.0 family, and a 3.3 billion parameter variant for 2.1. The 2.5 release scaled the shape model to roughly 10 billion parameters and lifted the output resolution to 1,024 cubed voxels. Hunyuan 3D 3.0 reaches 1,536 cubed, which Tencent describes as 3.6 billion voxels, and uses a coarse to fine sampling schedule that first blocks out gross form and then refines micro detail.
The texture model is named Hunyuan3D Paint. It takes the mesh from the shape stage and renders consistent textures across multiple viewpoints, then bakes those views into UV space. Versions 1.0 and 2.0 produced albedo only RGB textures. Version 2.1 introduced physically based rendering outputs, generating base colour, metallic, roughness, normal, and ambient occlusion maps in a single pass at up to 4K resolution. The 2.1 paint model reports 2 billion parameters, while the shape model reports 3.3 billion. A separate Hunyuan3D Delight model handles relighting and shading correction.
The original Hunyuan3D outputs were marching cubes meshes, which produced dense but artistically messy triangle soup. To address this the team released Hunyuan3D PolyGen in July 2025. PolyGen treats meshes as token sequences and generates them autoregressively using a custom Blocked and Patchified Tokenization scheme that the team claims compresses tokens per face by roughly 74 percent. The model supports both triangle and quadrilateral topology and can generate meshes with more than 20,000 faces while preserving clean edge flow suitable for animation and rigging. Tencent reports that internal artists using PolyGen reduced average modelling time from eight hours to two and a half hours per asset.
Hunyuan3D Omni, announced in September 2025, sits on top of the 2.1 base and accepts four additional conditioning signals: skeletal pose, sparse or dense point clouds, voxel grids, and 3D bounding boxes. Tencent describes it as the ControlNet of 3D. The system uses a lightweight unified control encoder rather than a separate head per modality, and a difficulty aware training curriculum that progressively introduces harder conditioning combinations.
The HunyuanWorld branch generates whole scenes rather than single objects. HunyuanWorld 1.0 produces a 360 degree panoramic proxy of a scene from a sentence or an image, then converts it into an explorable mesh with disentangled object layers. The output is compatible with standard computer graphics pipelines and is intended for virtual reality and game development. A lighter HunyuanWorld 1.0 Lite variant followed for resource constrained deployments, and HunyuanWorld 1.5 in December 2025 added real time interactive playback under the name WorldPlay. HunyuanWorld 2.0 shipped in April 2026.
The table below summarises the input modes, outputs, and feature additions reported across the major versions.
| Capability | 1.0 | 2.0 | 2.1 | 2.5 | 3.0 |
|---|---|---|---|---|---|
| Text to 3D | Yes | Yes | Yes | Yes | Yes |
| Image to 3D | Yes | Yes | Yes | Yes | Yes |
| Multi view image input | No | Partial | Yes | Yes | Yes (up to four views) |
| Sketch to 3D | No | No | Limited | Yes | Yes |
| PBR texture output | No | No | Yes | Yes (4K) | Yes (4K with bump) |
| Geometric resolution | n/a | 256 cubed | 512 cubed | 1,024 cubed | 1,536 cubed |
| Quad mesh topology | No | No | Via PolyGen | Via PolyGen | Native and PolyGen |
| Skeletal rigging | No | No | No | Optimized skinning | Yes via Hunyuan 3D Studio |
| Output formats | OBJ | OBJ, GLB | OBJ, GLB | OBJ, GLB | OBJ, GLB |
| Engine compatibility | Generic | Blender addon | Blender, Unity | Blender, Unity, Unreal | Blender, Unity, Unreal |
| End to end production pipeline | No | No | No | Partial | Yes (Studio) |
Tencent reports that Hunyuan 3D 3.0 triples the geometric precision of its predecessor and reaches a CLIP score of 0.821 on Tencent's internal image to 3D benchmark, with what the company describes as a 15 percent improvement in geometric precision and a 20 percent improvement in texture fidelity over 2.5. These numbers come from Tencent's own evaluations and have not been independently reproduced at the time of writing.
Hunyuan3D weights are not released under a standard permissive licence. Instead Tencent uses a bespoke Tencent Hunyuan 3D Community License Agreement, first published with version 2.0 on 21 January 2025 and updated for 2.1 in June 2025. The licence grants a non exclusive, non transferable, royalty free right to use, reproduce, modify, and distribute the materials within a defined territory. That territory is the world, with the express exclusion of the European Union, the United Kingdom, and South Korea. Users in those three jurisdictions are not licensed to use, modify, or distribute the model under the community licence.
The agreement also restricts the use of model outputs to improve other AI models, prohibits military applications, and includes the usual indemnity and disclaimer clauses. Above a 100 million monthly active user threshold, commercial users must request a separate licence from Tencent. The territory restriction is the most contentious element. Community threads on GitHub and Hacker News in early 2025 pressed the team to switch to Apache 2.0 or MIT, but Tencent has so far kept the carve out, citing regulatory uncertainty in the excluded regions.
Not every component of the project ships under the same terms. The Dust3R based baking module in Hunyuan3D 1.0 was provided under Creative Commons BY NC SA 4.0, which is non commercial. Subsequent releases avoided that dependency. Hunyuan3D 2.1 was the first version to release the full training code in addition to the weights, which Tencent describes as fully open source within the licensed territory.
The global Hunyuan 3D creation engine is offered through two channels. The first is a consumer web platform at the Hunyuan 3D portal, which gives every signed in user 20 free generations per day. The second is an enterprise API on Tencent Cloud, available in English and aimed at integration into existing applications. New enterprise accounts receive 200 free credits at registration.
Beyond the free tier the API uses a prepaid credit model. A Professional Edition call that generates a 3D model from text or image with ordinary texture mode consumes 25 credits, with additional options raising the cost. An Express Edition call in the same mode consumes 15 credits. Credit packs are valid for one year from purchase and unused credits expire at the end of that window. The settlement order applies prepaid packs before falling back to pay as you go billing. Tencent has not published a fixed USD per credit rate in English, leaving regional list pricing to the Tencent Cloud console.
Output formats include OBJ and GLB, and Tencent advertises direct compatibility with Unity, Unreal Engine, and Blender. Specific platform integrations include an official Blender addon released in January 2025 and a Tencent maintained ComfyUI workflow for Hunyuan 3D 3.0 published in late 2025.
The table below compares Hunyuan 3D 3.0 with three commonly cited rivals in the generative 3D space as of May 2026. Figures come from each vendor's public documentation and from independent reviews where cited; capabilities not confirmed by primary sources are left blank.
| System | Vendor | Open weights | PBR textures | Geometric resolution | Notable strength |
|---|---|---|---|---|---|
| Hunyuan 3D 3.0 | Tencent | Yes, community licence | Yes, 4K | 1,536 cubed | Hard surface geometry, scale, free open weights |
| Tripo P1 | Tripo AI | No | Yes | Not disclosed | Organic shapes, generation speed |
| Meshy 6 | Meshy | No | Yes | Not disclosed | Stylized character art, mature web product |
| Rodin Gen-2 | Deemos | No | Yes | Not disclosed | High end realism, sculpt quality |
In head to head reviews from outlets such as 3D AI Studio, Vset3D, and the Scenario knowledge base, reviewers tend to converge on a pattern. Hunyuan is favoured for hard surface modelling and for the price of admission, since the weights are free to download. Tripo is praised for organic creatures and stylised content and for fast iteration. Meshy is rated highly for its product polish, documentation, and security certifications including SOC 2 and ISO 27001. Rodin Gen 2 is positioned as the realism leader for high budget productions. These assessments are inevitably subjective and depend heavily on the prompts and reference images used.
The response to Hunyuan 3D inside the open source community has been broadly positive. The three million Hugging Face downloads reported at the global launch make it the most downloaded 3D foundation model on the platform, ahead of Stable Zero123 and Trellis. ComfyUI maintainers integrated Hunyuan 3D 3.0 nodes within weeks of the September 2025 release, and the Blender addon has become a standard component of indie 3D pipelines.
Criticism has focused on three areas. The first is the territory carve out in the community licence, which prevents legal commercial use in the European Union, the United Kingdom, and South Korea and frustrates contributors based in those regions. The second is hardware cost. Running the 2.1 combined pipeline requires roughly 29 GB of VRAM, and the 2.5 and 3.0 models push that further, putting native inference out of reach of consumer GPUs without quantization. A community fork named Hunyuan3D 2GP appeared to address low VRAM environments. The third is mesh quality at the low end. Reviewers consistently noted that the raw marching cubes output before PolyGen was unsuitable for animation, and that even with PolyGen the topology still needs cleanup for facial rigs.
Academic reception has tracked the open releases closely. The HunyuanWorld 1.0 technical report appeared on arXiv as 2507.21809 in July 2025 and the Hunyuan 3D Studio paper followed on arXiv in September 2025 as 2509.12815. Both have been cited in subsequent work on 3D scene generation and end to end 3D asset pipelines.
Industrial adoption has been led by Tencent's own properties, including Tencent Games and Tencent Music, and by partner integrations announced at the global launch. Bambu Lab integrated Hunyuan 3D as an input source for its consumer 3D printers, allowing users to print AI generated models directly. Unity China shipped a Hunyuan 3D plugin for the Unity editor. Liblib added Hunyuan 3D as a generation option alongside its existing image models. Outside Tencent's direct partnerships, the model has also been picked up by independent 3D printing services, e commerce visualization startups, and indie game developers who appreciate the open weights and the absence of per asset licensing fees within the permitted territory.