BrushGaussian: Brushstroke-Based Stylization for 3D Gaussian Splatting

BrushGaussian is a method for enhancing 3D Gaussian Splatting with brushstroke-aware stylization — supporting both color transfer and artistic shape deformation. Unlike previous 3D style transfer methods that only adjust color or texture, we enable fine-grained control over geometry, inspired by real brushstrokes.

✨ Our method:

• Adds per-Gaussian texture features to simulate brushstroke-like effects;

• Applies a custom texture mapping technique for geometric stylization;

• Uses unsupervised clustering to prune redundant Gaussians for efficiency;

• Is fully compatible with existing 3DGS pipelines.

Cloning

git clone https://github.com/Xzzit/BrushGS --recursive
cd BrushGS

Installation

There are two ways to set up the environment depending on whether you've already installed 3D Gaussian Splatting.

Option 1: You already have 3DGS installed

conda activate <your_3DGS_env>
pip install scikit-learn==1.0.2
pip install brush-gaussian-rasterization\.

Option 2: You don't have 3DGS installed

Create a new conda environment with all required packages:

conda env create -f environment.yml

💡Tip: The default environment name is brushGS, but feel free to change it in environment.yml if needed.

Running

Step 0: (Optional) Prepare a 3D Gaussian Splatting Model

If you don't already have a pre-trained 3DGS model, follow the instructions in the 3D Gaussian Splatting repository to train a Gaussian Splatting model on your dataset.

After training, the output directory of your model should look like this:

name_of_the_scene/
├── point_cloud/               # Contains Gaussian checkpoints
│   ├── iteration_7000/
│   │   └── point_cloud.ply
│   ├── iteration_30000/
│   └── ...
├── cameras.json
├── exposure.json
├── cfg_args
├── input.ply
├── train/                     # Optional
├── test/                      # Optional

Step 1: Pruning and Clustering

Run the following command to prune the original Gaussians and perform clustering:

python pruning.py -m <path_to_your_model_name> -n <number_of_clusters>

Example:

python pruning.py -m E:\gaussian-splatting\output\mic -n 250

Arguments:

-m <path_to_your_model_name>: Path to the trained 3DGS model directory.

-n <number_of_clusters>: Number of output clusters (i.e., the number of retained Gaussian primitives). This value depends on the complexity of your scene:

• 250 ~ 1000 for a simple object scene (e.g., NeRF synthetic)
• 7000 ~ for complex scenes (e.g., Tanks and Temples)

Step 2: Stylized Rendering

Render the pruned Gaussian model using a custom brushstroke-style image:

python render.py -m <path_to_your_model> -t <style_image_path>

Example:

python render.py -m output\mic -t texture\splatting.png

Arguments:

-m <path_to_your_model>: Path to the pruned Gaussian model directory.

-t <style_image_path>: Path to the style image you want to apply. For now, we only support RGBA images with alpha channels. If not provided, a elliptical rendering will be performed.

-n <number_of_cluster>: Which cluster to be rendered. Default is -1, stands for max cluster.

Known Issues

• The current 3DGS has a bug when training on nerf synthetic datasets, which can be found in these issue1 and issue2.