Hunyuan3D-PolyGen - A large-scale, art-grade 3D model generated by Tencent Hunyuan.

What is Hunyuan3D-PolyGen?

Hunyuan3D-PolyGen is the industry's first art-grade 3D model generator, launched by Tencent's Hunyuan team. Combining its self-developed high-compression-ratio representation technology (BPT), it can generate complex geometric models with tens of thousands of faces, supporting both triangular and quadrilateral faces to meet the needs of different professional pipelines. The model employs an autoregressive mesh generation framework, generating high-quality, art-compliant 3D models through three steps: mesh serialization, autoregressive modeling, and sequence decoding. A reinforcement learning post-training framework is introduced to improve the stability and artistic accuracy of the generated models. Hunyuan3D-PolyGen primarily solves the challenges of topology quality and complex object modeling in 3D asset generation, significantly improving the modeling efficiency of artists and applicable to UGC game asset generation.

Main functions of Hunyuan3D-PolyGen

• High-quality 3D model generationIt can generate complex geometric models with tens of thousands of faces, supporting three-sided and four-sided faces to meet the needs of different professional pipelines.
• Improve modeling efficiencyAfter being integrated into multiple Tencent game pipelines, the modeling efficiency of artists has increased by over 70%.
• Adaptive face number allocationIt can adaptively allocate the number of faces based on the geometric structure, achieving better detail representation under low face count conditions.
• Reinforcement learning optimizationBy using reinforcement learning to train the framework, the stability and aesthetic consistency of the generative model can be improved.
• Multiple input methodsIt supports generating 3D models via text prompts or uploaded images, and users can also adjust the generation parameters to meet specific needs.

The technical principle of Hunyuan3D-PolyGen

• Autoregressive Mesh Generation Framework
  • Grid serializationThis function converts the vertices and faces of a 3D mesh into a sequence of tokens to represent the mesh structure. This enables the model to handle complex 3D mesh data in a sequential modeling manner.
  • Autoregressive modelingUsing point cloud as input prompt, a sequence of tokens for the mesh is generated based on an autoregressive model. The autoregressive model gradually constructs a complete 3D mesh structure by predicting tokens one by one.
  • Sequence DecodingThe generated token sequence is decoded in reverse into vertices and faces to reconstruct a 3D mesh. This ensures the geometric integrity and accuracy of the generated mesh.
• High compressibility characterization technique (BPT)To address the redundancy issue in existing mesh autoregressive methods, Hunyuan3D-PolyGen developed its own high-compression mesh representation technology, BPT (Blocked and Patchified Tokenization). By using block indexing and patch compression, the number of tokens required to represent the same mesh is significantly reduced, improving the model's ability to model complex meshes.
• Reinforcement learning post-training frameworkTo improve the stability of mesh autoregressive generation, Hunyuan3D-PolyGen introduces a reinforcement learning post-training framework. Post-training is performed on top of the pre-trained model, guiding it to generate higher-quality results through stable generation and art-standard rewards. Reinforcement learning effectively increases the probability of generating "good results" and reduces the probability of generating "bad results," significantly improving the stability of model generation.
• Transformer architectureThe autoregressive modeling part of Hunyuan3D-PolyGen adopts the Transformer architecture. The Transformer can capture long-range dependencies in mesh data and handle complex relationships between vertices and faces through the attention mechanism. The model is divided into two parts: a vertex model and a face model. The vertex model is responsible for generating the vertex sequence of the mesh, while the face model generates the face sequence of the mesh given the vertex sequence.

How to use Hunyuan3D-PolyGen

• Access the creation platformVisit the official platform of Tencent Hunyuan 3D AI Creation Engine.
• Function Entry:
  • Select the "3D Smart Topology" function on the homepage.
  • The "3D Smart Topology" function in the laboratory.
  • On the homepage, you can select "Geometric and Texture Stages" under the "Texture-to-Depth" or "Graphic-to-Depth" function. After generating the geometric model, you can select "Intelligent Topology".
• How to use
  • Vincent 3D:
    • Select the "Wensheng 3D" function on the homepage.
    • Enter text prompts to describe the features of the 3D model you want to generate.
    • Click "Generate" and wait for the model to be generated.
  • 3D graphics:
    • Select the "Image-to-3D" function on the homepage.
    • Upload an image (a non-transparent background is recommended).
    • Click "Generate," and the model will generate a corresponding 3D model based on the image.
  • 3D Intelligent Topology:
    • Select the "3D Smart Topology" function on the homepage or in the lab.
    • Upload a 3D model file (e.g., .obj format).
    • The model automatically optimizes the topology to generate higher quality 3D meshes.
• Adjust parametersDuring the generation process, you can adjust the following parameters:
  • Number of facesAdjust the number of faces in the generated model according to requirements.
  • Texture details: Choose whether to generate a texture map.
  • GIF animationChoose whether to generate a GIF animation to show the rotation effect of the 3D model.
• View and download resultsOnce generated, you can view the resulting 3D model. If needed, you can download the generated model file (e.g., .obj format) or GIF animation.

Application Scenarios of Hunyuan3D-PolyGen

• UGC game asset generationHunyuan3D-PolyGen can quickly generate high-quality 3D models, suitable for UGC (user-generated content) game development scenarios. Artists can use these models to quickly generate complex geometric models, significantly improving modeling efficiency by over 70%.
• Game scene and character modelingThe model supports the generation of high-resolution, highly detailed 3D models, which can meet the high requirements for scene and character modeling in game development.
• Virtual scene and prop generationHunyuan3D-PolyGen can generate complex geometric models, supporting three-sided and four-sided surfaces, meeting the high-precision requirements for virtual scenes and props in film and television production.
• Animation ProductionBy generating high-quality 3D models, these models can provide efficient modeling support for animation production, reducing the time and cost of manual modeling.
• Product Prototype DesignHunyuan3D-PolyGen can quickly generate high-fidelity 3D models, suitable for product prototyping in industrial design, helping designers quickly verify design concepts.
• Virtual assembly and displayThe generated 3D model can be directly used for virtual assembly and display, improving design communication efficiency.