Our implementation is based on the original 3D Gaussian splatting implementation available here.

Note that we implemented our method based on a former version of this software, so for consistence we recommend to use the same version.

Installation

Download 3D Gaussian Splatting

wget https://github.com/graphdeco-inria/gaussian-splatting/archive/d9fad7b3450bf4bd29316315032d57157e23a515.zip
unzip d9fad7b3450bf4bd29316315032d57157e23a515.zip
mv gaussian-splatting-d9fad7b3450bf4bd29316315032d57157e23a515  SGSST

We have used the style transfer loss from SPST: Scaling Paiting Style Transfer which should be cloned within the SGSST folder

cd SGSST
git clone https://github.com/bgalerne/scaling_painting_style_transfer.git

Download the original VGG19 weights vgg_conv.pth from here and save them in the scaling_painting_style_transfer/model/folder.

E.g. using gdown:

import gdown
gdown.download("https://drive.google.com/uc?id=1lLSi8BXd_9EtudRbIwxvmTQ3Ms-Qh6C8", "model/vgg_conv.pth")

Copy our stylization script and environment setting into SGSST/

cp  stylize.py  environment.yml  SGSST/

Local Setup

Our default installation method is based on Conda package and environment management:

conda env create --file environment.yml
conda activate SGSST

Download data

These datasets can be downloaded following the instruction of ARF, 3D gaussian splatting, and Plenoxels :

The SfM data sets from 3D gaussian splatting for Tanks&Temples and Deep Blending can be downloaded here.

The MipNeRF360 scenes are hosted by the paper authors here. You can put all of these data into the datasets folder.

Run SGSST:

Stylizing a scene is a two step procedure: First train a realistic 3DGS, then stylize it with our code.

• Run 3DGS reconstruction for the scene.

python train.py --source_path <path to COLMAP or NeRF Synthetic dataset>  \
                --model_path <path of the 3DGS output model> \
                --resolution 1  

Example:

python train.py  --source_path ./datasets/truck \
                 --model_path  ./output/model_truck \
                 --iterations 30000 \
                 --checkpoint_iterations 30000 \
                 --resolution 1  

• Stylize the scene with a given style image

python stylize.py --source_path <path to COLMAP or NeRF Synthetic dataset>  \
                  --model_path <path of the stylized 3DGS output model>     \
                  --start_checkpoint <path of the 3DGS input model>         \
                  --style_img  <path of stylized image>                     \
                  --iterations 50000 --resolution 1  

Example:

python stylize.py --source_path ./datasets/truck  \
                  --model_path output/model_truck_stylized  \
                  --start_checkpoint ./output/model_truck/chkpnt30000.pth \
                  --style_img  datasets/styles/112.jpg \
                  --iterations 50000 --resolution 1  

• Render the stylized scene

python render.py -m <path to the stylized 3DGS model> \
                 --source_path  <path to COLMAP or NeRF Synthetic dataset>

Example:

python render.py -m output/model_truck_stylized \
                  --source_path  ./datasets/truck

• generate a movie from stylized scene

python make_rendering_videos.py  -m <path to the stylized 3DGS model> 

Example:

python make_rendering_videos.py  -m output/model_truck_stylized --video