Image Motion Blur Removal in the Temporal Dimension with Video Diffusion Models

[Website] [arXiv] [Hugging Face]

Installation

conda create -n motion_deblur python=3.9
conda activate motion_deblur

# cuda-11.8 + torch-2.2.2
pip install torch==2.2.2 torchvision==0.17.2 torchaudio==2.2.2 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

Inference

To run inference with our provided weights, first download weights into folder pretrained as shown from Hugging Face.

pretrained/
├── bair
│   ├── ema.pt
│   └── vqgan.ckpt
└── clev
    ├── ema.pt
    └── vqgan.ckpt

For BAIR datasets,

python sample_motion_deblur.py configs/bair/deblur.py --ckpt-path pretrained/bair/ema.pt

For CLEVRER datasets,

python sample_motion_deblur.py configs/clev/deblur.py --ckpt-path pretrained/clev/ema.pt

Acknowledgement

• DiT: Scalable Diffusion Models with Transformers.
• DPS: Diffusion Posterior Sampling for General Noisy Inverse Problems.
• Open-Sora: Democratizing Efficient Video Production for All.
• VQGAN: Taming Transformers for High-Resolution Image Synthesis.

Citation

@INPROCEEDINGS{11084505,
  author={Pang, Wang and Zhan, Zhihao and Zhu, Xiang and Bai, Yechao},
  booktitle={2025 IEEE International Conference on Image Processing (ICIP)}, 
  title={Image Motion Blur Removal In The Temporal Dimension With Video Diffusion Models}, 
  year={2025},
  volume={},
  number={},
  pages={325-330},
  keywords={Deblurring;Visualization;Technological innovation;Dynamics;Estimation;Training data;Transformer cores;Diffusion models;Transformers;Kernel;Motion deblurring;video diffusion model;diffusion transformer},
  doi={10.1109/ICIP55913.2025.11084505}}