Hunyuan3D 2.0: High-Resolution 3D Assets Generation

Introduction

Hunyuan3D 2.0 represents a significant advancement in large-scale 3D synthesis systems, focusing on the generation of high-resolution textured 3D assets. This cutting-edge system comprises two primary components: Hunyuan3D-DiT, a shape generation model, and Hunyuan3D-Paint, a texture synthesis model. These components work in unison to produce intricate 3D geometries and vibrant texture maps, catering to both generated and handcrafted meshes.

Key Features

Advanced Architecture

The architecture of Hunyuan3D 2.0 revolves around a two-stage generation pipeline. This approach initiates with the formation of a bare mesh followed by texture map synthesis. Such a bifurcation allows for addressing the intricacies of shape and texture generation independently while offering flexibility for texturing options.

Performance and Evaluation

Hunyuan3D 2.0’s performance has been extensively evaluated against both open-source and closed-source 3D generation methodologies. The results demonstrate that Hunyuan3D 2.0 outperforms existing models in terms of the quality of generated 3D assets and alignment with specified conditions.

Model CMMD FID_CLIP FID CLIP-score
Hunyuan3D 2.0 3.193 49.165 282.429 0.809
Top Open-source Model1 3.591 54.639 289.287 0.787
Top Close-source Model1 3.600 55.866 305.922 0.779
Top Close-source Model2 3.368 49.744 294.628 0.806
Top Close-source Model3 3.218 51.574 295.691 0.799

Getting Started

Users can access Hunyuan3D 2.0 via code or by utilizing the Gradio App. The installation process involves setting up necessary requirements such as Pytorch and other dependencies. Utilizing the API, users can generate meshes with the Hunyuan3D-DiT pipeline and enhance them with textures through the Hunyuan3D-Paint pipeline.

Open-Source Initiatives

Hunyuan3D 2.0 is committed to open-source development, providing inference code, model checkpoints, and technical reports to the community. The team encourages collaboration and exploration within the 3D asset creation domain.

Acknowledgements

The development team acknowledges contributions from projects like DINOv2, Stable Diffusion, and others, which have been instrumental in driving innovation in 3D asset generation.