[TPAMI'25] MoBluRF: Motion Deblurring Neural Radiance Fields for Blurry Monocular Video

Minh-Quan Viet Bui^1*, Jongmin Park^1*, Jihyong Oh^2†, Munchurl Kim^1†
1KAIST, South Korea, ²Chung-Ang University, South Korea
*Co-first authors (equal contribution), †Co-corresponding authors

📣 News

Updates

• October 24, 2025: Code released.
• May 28, 2025: MoBluRF was published in TPAMI 2025 🎉.
• December 21, 2023: Paper uploaded to arXiv. Check out the manuscript here.

To-Dos

• Add Stereo Blur dataset configurations.

⚙️ Environment Setup

Tested on Python 3.8, PyTorch 1.13.1, CUDA 11.7, JAX 0.4.1.

git clone --recursive https://github.com/KAIST-VICLab/moblurf.git
cd moblurf

# Conda env
conda create -n moblurf python=3.8 -y
conda activate moblurf

# PyTorch (CUDA 11.7)
conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 pytorch-cuda=11.7 -c pytorch -c nvidia

# JAX (CUDA11)
python -m pip install --upgrade pip
pip install "jax==0.4.1" "jaxlib==0.4.1+cuda11.cudnn82" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html

# Project deps (no torch/jax in this file)
pip install -r requirements.txt

📁 Data Preparations

Blurry iPhone Dataset

We host the Blurry iPhone Dataset on Dropbox and the 🤗 Huggingface.
You can download our dataset via command-line below:

# Create root directory
mkdir -p data

# Download dataset zip from Dropbox
wget "https://www.dropbox.com/scl/fi/cez7kds7yoc6kdcsptrpq/blurry_iphone.zip?rlkey=nhdxuzsyfubqpliodksnue5xf&st=yfz33ums&dl=1" \
    -O data/blurry_iphone.zip

# Unzip
unzip data/blurry_iphone.zip -d data/blurry_iphone

# Or download on Huggingface
git clone https://huggingface.co/datasets/ImJongminPark/BlurryIPhone

After extraction, the dataset will follow this structure (example for apple):

data/blurry_iphone/
  ├── apple/
  │   ├── blurry_rgb/2x/*.png
  │   ├── normal/2x/*.npy
  │   ├── depth/2x/*.npy
  │   ├── camera_interp/*.json
  │   ├── covisible/2x/val/*.png
  │   ├── keypoint/2x/train/*.json
  │   ├── scene.json
  │   ├── dataset.json
  │   ├── metadata.json
  │   ├── extra.json
  │   ├── emf.json
  │   └── splits/

Training

Order matters. Train BRI first, then start MDD initialized from the BRI checkpoint.

1) BRI Stage (train first)

CUDA_VISIBLE_DEVICES=0,1,2,3 python tools/launch.py --gin_configs configs/experiments/blurry_iphone/runs/bri/apple.gin --gin_bindings Config.engine_cls=@Trainer --gin_bindings "SEQUENCE='apple'"

2) MDD Stage (initialize from the BRI checkpoint)

CUDA_VISIBLE_DEVICES=0,1,2,3 python tools/launch.py --gin_configs configs/experiments/blurry_iphone/runs/mdd/apple.gin --gin_bindings Config.engine_cls=@Trainer --gin_bindings "SEQUENCE='apple'"

Evaluation

python tools/launch.py --gin_configs configs/experiments/blurry_iphone/runs/mdd/apple.gin --gin_bindings Config.engine_cls=@Evaluator --gin_bindings "SEQUENCE='apple'"

⭐ Citing MoBluRF

If you find our repository useful, please consider giving it a star ⭐ and citing our research papers in your work:

@article{bui2025moblurf,
  title={Moblurf: Motion deblurring neural radiance fields for blurry monocular video},
  author={Bui, Minh-Quan Viet and Park, Jongmin and Oh, Jihyong and Kim, Munchurl},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2025},
  publisher={IEEE}
}

Acknowledgments

This project is developed with Dycheck codebase. We thank the original authors for their excellent work.