Required packages
November 5, 2025 · View on GitHub
[ICCV 2025] Stable Score Distillation
Formulating 2D image and 3D scene editing within diffusion models from the perspective of the score.
Contents
Installation
Our environment was tested on Ubuntu 22, CUDA 11.7 with 3090.
conda create -n ssd python=3.8 -y
conda activate ssd
conda install -c "nvidia/label/cuda-11.7.0" cuda-nvcc
conda install cuda-toolkit==11.7
pip install ninja
pip install cmake
conda install pytorch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 pytorch-cuda=11.7 -c pytorch -c nvidia
# Gaussian Splatting
cd gaussiansplatting
pip install submodules/diff-gaussian-rasterization
pip install submodules/simple-knn
# Required packages
cd ..
pip install tqdm
pip install plyfile
pip install mediapipe
pip install diffusers==0.27.2
pip install -r requirements_all.txt
We provide an environment.yaml file to help you verify.
Tips
-
If the default resolution of 512×512 is not suitable, you can modify the line here by setting
self.use_original_resolutiontoTrue, and adjust the resolution accordingly. For example, if the original(height, width)is(729, 985), you may change it to something like(512, 692). -
Some prompts may not work well with SD2.1. In such cases, you can try using IP2P instead.
-
Most configurations are adapted from GaussianEditor.
Our pipeline relies on three key configuration:cross-prompt,cross-trajectory, andprompt-enhancement.
If the default values — cross-trajectory: 2.0 and enhance_scale: 5.5 — lead to suboptimal results, users can try adjusting the weights. -
We provide test data and the evaluation metric code.
Command Line
Please try our demo by running script/face.sh.
Citation
If you find our work helpful in your project, please cite:
@InProceedings{Zhu_2025_ICCV,
author = {Zhu, Haiming and Xu, Yangyang and Xu, Chenshu and Shen, Tingrui and Liu, Wenxi and Du, Yong and Yu, Jun and He, Shengfeng},
title = {Stable Score Distillation},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
month = {October},
year = {2025},
pages = {16597-16606}
}
Acknowledgement
Most of our code is adapted from the excellent works of GaussianEditor and Threestudio. We sincerely thank the authors for their great contributions.
We also refer to the following projects: