StyleSculptor: Zero-Shot Style-Controllable 3D Asset Generation with Texture-Geometry Dual Guidance

Zefan Qu · Zhenwei Wang · Haoyuan Wang · Ke Xu · Gerhard Hancke · Rynson W.H. Lau

City University of Hong Kong

SIGGRAPH Asia 2025

📢 News

• [Aug.10.2025] StyleSculptor is accepted to SIGGRAPH Asia 2025 🍀. The code is still being organized. Stay tuned！
• [Oct.15.2025] We release the inference code of dual-style guidance generation.

📦 Installation

Our code is highly built on the TRELLIS Repo, you can follow their offical guidance and find the solutions of the installation problems.

Prerequisites

• System: The code is currently tested only on Linux.
• Hardware: An NVIDIA GPU with at least 16GB of memory is necessary. The inference time of each case is about 2 minutes on a NVIDIA RTX 4090.
• Software:
  • The CUDA Toolkit is needed to compile certain submodules.
  • Conda is recommended for managing dependencies.
  • Python version 3.8 or higher is required.

Installation Steps

1. Clone the repo:

   git clone https://github.com/quzefan/StyleSculptor
   cd StyleSculptor

2. Install the dependencies: Create a new conda environment named stylesculptor and install the dependencies:

   . ./setup.sh --new-env --basic --xformers --flash-attn --diffoctreerast --spconv --mipgaussian --kaolin --nvdiffrast

🤖 Pretrained Models

We don't modify the pretrained models of TRELLIS. You can directly load the models with their repository names in the code:

TrellisImageTo3DPipeline.from_pretrained("microsoft/TRELLIS-image-large")

You can also load the model from local by downloading the checkpoint (TRELLIS-image-large) from TRELLIS Repo, then change the code like:

TrellisImageTo3DPipeline.from_pretrained("/path/to/TRELLIS-image-large")

For more information, please refer to TRELLIS.

💡 Inference

Data Preparation

The edge map of each style images should be provided. The format can be referred in ./asset/style_image_edge folder.

We use PidiNet to generate the edge map for style images.

⭐⭐Make sure there is main object in the content and style images. If not, please turn off the rembg operation in the data preprocessing stage.

Quick Start

python example.py --cnt /path/cnt_image --sty /path/sty_image --sty_edge /path/sty_edge_image --intensity intensity_value

Parameters in the command:

• cnt: Content image(s). Image path / Multi-view images Folder path.
• sty: Style image(s). Image path / Multi-view images Folder path.
• sty_edge: Edge maps of all input style images.
• intensity: The style guidance intensity. Valid value: 0 (No Guidance) - 5 (Full Guidance).

Example

python example.py --cnt ./asset/content_multi_image/character --sty ./asset/style_image/groot_010.png --sty_edge ./asset/style_image_edge/groot_010.png --intensity 2

📜 Citation

If you find this work helpful, please consider citing our paper:

@article{qu2025stylesculptor,
  title={StyleSculptor: Zero-Shot Style-Controllable 3D Asset Generation with Texture-Geometry Dual Guidance},
  author={Qu, Zefan and Wang, Zhenwei and Wang, Haoyuan and Xu, Ke and Hancke, Gerhard and Lau, Rynson WH},
  journal={arXiv preprint arXiv:2509.13301},
  year={2025}
}

Acknowledgements

Our codebase builds on TRELLIS. Thanks the authors for sharing their awesome codebases!