Official implementation of the paper titled "Scene-level Appearance Transfer with Semantic Correspondences".

Liyuan Zhu¹, Shengqu Cai^1,*, Shengyu Huang^2,*, Gordon Wetzstein¹, Naji Khosravan³, Iro Armeni¹

¹Stanford University, ²NVIDIA Research, ³Zillow Group | ^* denotes equal contribution

@inproceedings{zhu2025_restyle3d,
    author = {Liyuan Zhu and Shengqu Cai and Shengyu Huang and Gordon Wetzstein and Naji Khosravan and Iro Armeni},
    title = {Scene-level Appearance Transfer with Semantic Correspondences},
    booktitle = {ACM SIGGRAPH 2025 Conference Papers},
    year = {2025},
  }

We introduce ReStyle3D, a novel framework for scene-level appearance transfer from a single style image to a real-world scene represented by multiple views. This method combines explicit semantic correspondences with multi-view consistency to achieve precise and coherent stylization.

• Ubuntu 22.04 LTS, Python 3.10.15, CUDA 12.2, GeForce RTX 4090/3090

• CentOS Linux 7, Python 3.12.1, CUDA 12.4, NVIDIA A100

git clone [email protected]:GradientSpaces/ReStyle3D.git
cd ReStyle3D

conda create -n restyle3d python=3.10
conda activate restyle3d
pip install -r requirements.txt

Download the pretrained models by running:

bash scripts/download_weights.sh

We download our dataset:

bash scripts/download_data.sh

We include 3 demo images to run semantic appearance transfer:

python restyle_image.py

To run on a single scene and style:

python restyle_scene.py   \
 --scene_path demo/scene_transfer/bedroom_0/  \
 --scene_type bedroom   \
 --style_path demo/design_styles/bedroom/pexels-itsterrymag-2631746

We organize the data into two components:

1. Interior Scenes: Multi-view real-world scans with aligned images, depth, and semantic segmentations.

📁 data/
  └── interiors/
      ├── bedroom/
      │   ├── 0/
      │   │   ├── images/      # multi-view RGB images
      │   │   ├── depth/       # depth maps
      │   │   └── seg_dict/    # semantic segmentation dictionaries
      │   └── 1/
      │       └── ...
      ├── living_room/
      └── kitchen/

2. Design Styles: Style examplars with precomputed semantic segmentation.

📁 data/
  └── design_styles/
      ├── bedroom/
      │   └── pexels-itsterrymag-2631746/
      │       ├── image.jpg        # style reference image
      │       ├── seg_dict.pth     # semantic segmentation dictionary 
      │       └── seg.png          # segmentation visualization
      ├── living_room/
      └── kitchen/

• Release full dataset
• Release evaluation code
• Customize dataset

Our codebase is built on top of the following works:

• Cross-image-attention
• ODISE
• ViewCrafter
• GenWarp
• DUSt3R

We appreciate the open-source efforts from the authors.

If you encounter any issues or have questions, feel free to reach out: Liyuan Zhu.