Hunyuan3D 2.0 - AI Image Models Tool

Overview

Hunyuan3D 2.0 is a two-stage, diffusion-based system from Tencent for producing high-resolution textured 3D assets from images or text. The pipeline separates shape and texture: Hunyuan3D-DiT (a flow-matching diffusion transformer) generates a detailed mesh aligned to a conditioning image or text prompt, and Hunyuan3D-Paint synthesizes high-resolution UV texture maps that can be applied to generated or hand-crafted meshes. The project includes a production-facing UI (Hunyuan3D-Studio), a Blender addon, a Gradio demo and example integrations to help artists and developers iterate quickly. ([huggingface.co](https://huggingface.co/tencent/Hunyuan3D-2)) Hunyuan3D 2.0 was released publicly with model weights, inference code, and technical report (arXiv) so researchers and practitioners can run locally, extend the system, or integrate it into pipelines. The release also includes smaller and turbo/distilled variants (for lower VRAM or faster inference) and multiview-capable models. Community feedback has been broadly enthusiastic about output quality while noting practical caveats such as baked lighting in textures, topology limitations for rigging, and installation/VRAM needs. ([arxiv.org](https://arxiv.org/abs/2501.12202))

Model Statistics

• Downloads: 63,204
• Likes: 1691
• Pipeline: image-to-3d

License: other

Model Details

Architecture and components: Hunyuan3D 2.0 uses a two-stage architecture. Stage 1 is Hunyuan3D-DiT, a flow-matching diffusion transformer that generates octree-based triangle meshes conditioned on a single image, multiple views, or text. Stage 2 is Hunyuan3D-Paint, a diffusion-based texture synthesis model that produces high-resolution UV texture maps and (optionally) delighting / illumination compensation modules. The repository exposes diffusers-like pipelines (Hunyuan3DDiTFlowMatchingPipeline and Hunyuan3DPaintPipeline) and returns trimesh objects that can be exported to GLB/OBJ. ([huggingface.co](https://huggingface.co/tencent/Hunyuan3D-2)) Model sizes and variants: Tencent published multiple variants: the main DiT shape models are ~1.1B parameters, the Paint/texture model around ~1.3B parameters, and a 0.6B “mini” DiT variant for faster, lower-VRAM inference. Distilled/turbo releases and multiview finetunes (1.1B multiview models and step-distilled turbo variants) are provided to reduce latency or improve multiview consistency. VRAM and runtime: the official README lists ~6 GB VRAM for just shape generation and about 16 GB for the full shape-plus-texture pipeline, with mini/turbo variants requiring less. ([github.com](https://github.com/Tencent-Hunyuan/Hunyuan3D-2)) Outputs and integration: outputs are triangle meshes and textured UV maps (trimesh objects, exportable to GLB/OBJ). The codebase includes a Gradio app, a Blender addon, example scripts (minimal_demo.py), and references for integrating with ComfyUI / third-party wrappers. The project is open-source with a Tencent community license and published technical report (arXiv). ([github.com](https://github.com/Tencent-Hunyuan/Hunyuan3D-2))

Key Features

• Two-stage pipeline: DiT for geometry and Paint for high-resolution UV texture maps.
• Supports image-to-3D and text-to-3D workflows with diffusers-like API.
• Variants: 0.6B mini, 1.1B standard DiT, 1.3B Paint, and turbo/distilled releases.
• Production tools: Hunyuan3D‑Studio web demo, Blender addon, and Gradio example app.
• Open-source release with checkpoints, inference code, and technical report (arXiv).

Example Usage

Example (python):

from hy3dgen.shapegen import Hunyuan3DDiTFlowMatchingPipeline
from hy3dgen.texgen import Hunyuan3DPaintPipeline

# 1) Generate a mesh from an image (image-to-3D)
shape_pipe = Hunyuan3DDiTFlowMatchingPipeline.from_pretrained('tencent/Hunyuan3D-2')
mesh = shape_pipe(image='assets/reference.png', num_inference_steps=30)[0]

# mesh is a trimesh.Trimesh-like object; save as GLB/OBJ
mesh.export('output/generated_mesh.glb')

# 2) Run texture synthesis (apply high-res texture to the generated mesh)
paint_pipe = Hunyuan3DPaintPipeline.from_pretrained('tencent/Hunyuan3D-2')
textured_mesh = paint_pipe(mesh, image='assets/reference.png')
textured_mesh.export('output/generated_textured_mesh.glb')

# Notes: installation requires building the custom rasterizer and differentiable renderer per README.
# For text-to-3D use, see minimal_demo.py in the repository for prompt examples. ([huggingface.co](https://huggingface.co/tencent/Hunyuan3D-2))

Pricing

The Hunyuan3D 2.0 models and inference code are published open-source (Tencent community license) and the repository, Hugging Face model downloads, and technical report are freely available. Tencent also provides a web demo (Hunyuan3D-Studio) for quick testing. There is no official paid licensing or per-generation pricing published by Tencent for the model itself; however, third‑party services or commercial platforms that host the model may charge for hosted inference. For model weights, code and demos see the GitHub and Hugging Face pages. ([github.com](https://github.com/Tencent-Hunyuan/Hunyuan3D-2))

Benchmarks

CMMD (lower is better): 3.193 (Source: https://huggingface.co/tencent/Hunyuan3D-2)

FID_CLIP (lower is better): 49.165 (Source: https://huggingface.co/tencent/Hunyuan3D-2)

FID (lower is better): 282.429 (Source: https://huggingface.co/tencent/Hunyuan3D-2)

CLIP-score (higher is better): 0.809 (Source: https://huggingface.co/tencent/Hunyuan3D-2)

Hugging Face: downloads (last month): 63,204 (Source: https://huggingface.co/tencent/Hunyuan3D-2)

Hugging Face: likes: 1.69k (Source: https://huggingface.co/tencent/Hunyuan3D-2)

Key Information

• Category: Image Models
• Type: AI Image Models Tool