FantasyStyle: Controllable Stylized Distillation for 3D Gaussian Splatting

Yitong Yang¹,   Yinglin Wang^1†,   Changshuo Wang³,   Huajie Wang^4,5,   Shuting He^1,2†,  

¹School of Computing and Artificial Intelligence, Shanghai University of Finance and Economics, Shanghai, China.
²MoE Key Laboratory of Interdisciplinary Research of Computation and Economics, Shanghai, China.
³Department of Computer Science University College London, London, United Kingdom.
⁴Shandong University of Finance and Economics, Shandong, China.
⁵Jinan Yunwei Software Technology Co., Ltd, Shandong, China
^†Corresponding Author.

   

💡 Overview

We introduce FantasyStyle, a 3DGS-based style transfer framework, and the first to rely entirely on diffusion model distillation. It comprises two key components: (1) Multi-View Frequency Consistency. We enhance cross-view consistency by applying a 3D filter to multi-view noisy latent, selectively reducing low-frequency components to mitigate stylized prior conflicts. (2) Controllable Stylized Distillation. To suppress content leakage from style images, we introduce negative guidance to exclude undesired content. In addition, we identify the limitations of Score Distillation Sampling and Delta Denoising Score in 3D style transfer and remove the reconstruction term accordingly. Building on these insights, we propose a controllable stylized distillation that leverages negative guidance to more effectively optimize the 3D Gaussians.

🔧 Prepare

The repository contains the 3DGS project. Please follow the commands below to install it.

git clone https://github.com/graphdeco-inria/gaussian-splatting --recursive

You must install the environment required for 3D Gaussian Splatting. Then, follow the commands below to install our environment.

git clone https://github.com/yangyt46/FantasyStyle.git
cd FantasyStyle
pip install -r requirements.txt

📂 Datasets

IN2N

Tandt DB

Mip-NeRF 360

🚀 Run

Reconstruction scene

Reconstruct a scene based on 3DGS, with an example as follows:

python train.py -s <path to COLMAP or NeRF Synthetic dataset>

Style transfer

Perform style transfer based on the reconstructed scene, with an example as follows:

bash truch.sh

🎓 Citing FantasyStyle

If you use FantasyStyle in your research, please use the following BibTeX entry.

@article{yang2025fantasystyle,
  title={Fantasystyle: Controllable stylized distillation for 3d gaussian splatting},
  author={Yang, Yitong and Wang, Yinglin and Wang, Changshuo and Wang, Huajie and He, Shuting},
  journal={arXiv preprint arXiv:2508.08136},
  year={2025}
}