Arbitrary-Scale 3D Gaussian Super-Resolution

AAAI 2026

Huimin Zeng, Yue Bai, Yun Fu

Overview

Existing 3D Gaussian Splatting (3DGS) super-resolution methods typically perform high-resolution (HR) rendering of fixed scale factors, making them impractical for resource-limited scenarios. Directly rendering arbitrary-scale HR views with vanilla 3DGS introduces aliasing artifacts due to the lack of scale-aware rendering ability, while adding a post-processing upsampler for 3DGS complicates the framework and reduces rendering efficiency.

To tackle these issues, we build an integrated framework that incorporates scale-aware rendering, generative prior-guided optimization, and progressive super-resolving to enable 3D Gaussian super-resolution of arbitrary scale factors with a single 3D model. Notably, our approach supports both integer and non-integer scale rendering to provide more flexibility.

Key Features:

• ✅ Arbitrary-scale super-resolution (integer and non-integer scales)
• ✅ Single model for all scale factors
• ✅ Real-time rendering at high resolutions (85 FPS at 1080p)
• ✅ Significant quality improvement (6.59 dB PSNR gain over vanilla 3DGS)

🔥 News

• 2026/1/19: We released code and checkpoints for this project.

• 2026/1/16: Check out the project page for interactive comparisons!

• 2025/11/15: Our Arbi-3DGSR was accepted to AAAI 2026.

Preparation

Requirements

• Python 3.10
• CUDA 12.1 or later
• PyTorch 2.1.2 (with CUDA 12.1 support)
• torchvision 0.16.2
• pytorch-lightning 1.4.2

Option 1: Docker (Recommended)

We provide a Docker image environment for easy setup:

docker pull zeldam1/zhm_docker:zhm-py310-torch21
docker run --gpus all -it -v /workdir:/workdir  --shm-size 64g zeldam1/zhm_docker:zhm-py310-torch21 /bin/bash

cd Arbi-3DGSR

pip install pyiqa==0.1.10 pytorch-lightning==1.4.2 torchmetrics==0.6.0 taming-transformers-rom1504 scikit-learn kornia==0.6 open_clip_torch==2.0.2 transformers==4.38.2 clip accelerate==1.12.0 submodules/simple-knn submodules/diff-gaussian-rasterization

Option 2: Conda Environment

# Clone the repository
git clone git@github.com:huimin-zeng/Arbi-3DGSR.git
cd Arbi-3DGSR

# Create conda environment
conda create -y -n Arbi-3DGSR python=3.10
conda activate Arbi-3DGSR

# Install PyTorch with CUDA 12.1
pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu121


# Install other dependencies
pip install pytorch-lightning==1.4.2 torchmetrics==0.6.0 open-clip-torch==2.0.2
pip install pyiqa==0.1.10 taming-transformers-rom1504 scikit-learn kornia==0.6 transformers==4.38.2 clip accelerate==1.12.0

# Install submodules
pip install submodules/diff-gaussian-rasterization
pip install submodules/simple-knn

Dataset Preparation

Download datasets from Gaussian Splatting and prepare them with the following structure. Each scene should contain images at different resolutions for training and evaluation:

data/
├── db/
│   ├── drjohnson/
│   │   ├── images/          # Original resolution images
│   │   ├── images_x2/       # 2x downsampled images
│   │   ├── images_x4/       # 4x downsampled images
│   │   ├── images_x8/       # 8x downsampled images (training resolution)
│   │   ├── images_x8_3.5/   # Non-integer scale (8x -> 3.5x downsampled)
│   │   ├── images_x8_5.7/   # Non-integer scale (8x -> 5.7x downsampled)
│   │   └── sparse/          # COLMAP reconstruction
│   └── playroom/
│       └── ...              # Same structure as drjohnson
├── MipNeRF360/
│   ├── bicycle/
│   ├── flowers/
│   └── ...                  # Other scenes
├── synthetic_nerf_blender/
│   ├── chair/
│   ├── drums/
│   └── ...                  # Other scenes
└── tandt/
    ├── train/
    └── truck/

Note:

• To create multi-resolution datasets, you can use image downsampling scripts to generate images_x2, images_x4, images_x8 from the original images/ folder.

• Non-integer scale folders (images_x8_3.5/, images_x8_5.7/) are optional but recommended for arbitrary-scale evaluation

Quick Start

Check out the scripts in script/ folder, all scripts support the following flags:

• --train: Run training only
• --render: Run rendering only
• --eval: Run evaluation only
• No flags: Run all (train, render, eval)

Training

Train a model from scratch on the Deep Blending dataset:

# Train only
bash script/db.sh --train

# All-in-one: train, render, and evaluate
bash script/db.sh

Other datasets:

• NeRF Synthetic: bash script/blender.sh --train
• MipNeRF-360: bash script/mipnerf360.sh --train
• Tanks & Temples: bash script/tandt.sh --train

Rendering

1. We provide pretrained checkpoints at Google Drive

2. Extract and organize the checkpoints:

output/
├── db/
│   ├── drjohnson/
│   └── playroom/
├── synthetic_nerf_blender/
│   ├── chair/
│   ├── drums/
│   ├── ficus/
│   ├── hotdog/
│   ├── lego/
│   ├── materials/
│   ├── mic/
│   └── ship/
└── MipNeRF360/
    ├── bicycle/
    ├── flowers/
    └── ...

3. Render at different scales:

# Render only
bash script/db.sh --render

# Or render for other datasets
bash script/blender.sh --render
bash script/mipnerf360.sh --render

Evaluation

Compute image quality metrics (PSNR, SSIM, LPIPS):

bash script/db.sh --eval

Citation

If you find this work useful, please give us a star 🌟 and consider citing our paper ❤️:

@inproceedings{zeng2026arbitrary,
  title={Arbitrary-Scale 3D Gaussian Super-Resolution},
  author={Zeng, Huimin and Bai, Yue and Fu, Yun},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={40},
  number={15},
  pages={12304--12312},
  year={2026}
}

Acknowledgments

This repository builds upon excellent prior work:

• 3D Gaussian Splatting by Kerbl et al.
• Mip-Splatting by Yu et al.
• StableSR by Wang et al.

We thank their authors for sharing these excellent works!

License

This project is licensed under the MIT License.

Contact

For questions and issues, please open an issue on GitHub or contact zeng.huim@northeastern.edu.