Hunyuan3D 2.0

What Hunyuan3D 2.0 is

Hunyuan3D 2.0 is a large-scale 3D generation system for producing high-resolution 3D assets with textures. Given a condition image, it generates geometry that matches the image and then adds a texture map on top of it^[1]. Tencent released the inference code and the pretrained models on 21 January 2025, and the technical report was posted to arXiv the same day, so the paper, the code and the weights appeared together^[1][2].

Two models: shape and texture

The system is built around two main models. Hunyuan3D-DiT is a flow-based diffusion transformer that generates the shape geometry, aligned with the condition image^[1]. Diffusion models learn to turn noise into structure by removing noise step by step; a flow-based variant is trained to follow a smooth path from noise to data, and putting this inside a transformer lets the model work over the long sequences that a 3D shape amounts to. What comes out is geometry that matches what the input image shows^[1].

Hunyuan3D-Paint handles appearance: it synthesises high-resolution texture maps^[1]. It is not limited to shapes the system generated itself, producing high-resolution and vibrant texture maps for hand-crafted meshes as well^[1]. Splitting geometry from texture means each half can be trained and replaced on its own.

Sizes and hardware

The shape models have 1.1 billion parameters and the Paint texture models 1.3 billion^[2]. Running them is comparatively light: shape generation needs about 6 GB of VRAM, and shape plus texture about 16 GB^[2]. Those numbers decide who can actually use the system, since the shape stage fits on a single modest GPU.

Studio

The release also includes Hunyuan3D-Studio, a production platform for re-creating, manipulating and animating 3D assets^[1]. Where the models produce a static asset, the studio is where that asset is edited and set in motion.

Reported results

Tencent reported that the system beat the top open-source and closed-source baselines on four measures: CMMD at 3.193, FID_CLIP at 49.165, FID at 282.429 and CLIP-score at 0.809^[2]. The authors also claim it outperforms previous state-of-the-art open and closed models on geometry detail, alignment to the condition image and texture quality^[1]. These are the authors' own comparisons against the baselines they chose.

Where it sits in the line

Hunyuan3D 2.0 succeeds Hunyuan3D 1.0, whose technical report is arXiv 2411.02293^[2]. The line then moved quickly: later releases included Hunyuan3D-2mini, a 0.6-billion-parameter variant, and Hunyuan3D-2mv on 18 March 2025, Turbo variants on 19 March 2025, Hunyuan3D-2.1 with a PBR texture model on 13 June 2025, and the Hunyuan3D 2.5 technical report on 23 June 2025^[2]. That sequence of small, multi-view, faster and physically based variants shows the base system being pushed along several axes at once rather than replaced by a single follow-up.