Pluralistic Image Completion with Gaussian Mixture Models (PICMM)

Paper

This repository is the official pytorch implementation of our NeurIPS 2022 paper, Pluralistic Image Completion with Gaussian Mixture Models.

Xiaobo Xia^1*, Wenhao Yang^2*, Jie Ren³, Yewen Li⁴, Yibing Zhan⁵, Bo Han⁶, Tongliang Liu^1
1The University of Sydney, ²Nanjing University, ³The University of Edinburgh, ⁴Nanyang Technological University, ⁵JD Explore Academy, ⁶Hong Kong Baptist University
* Equal contributions

Prerequisites

• Python >=3.7
• NVIDIA GPU + CUDA cuDNN

pip install -r requirements.txt

Pipeline

Training

python train.py --name [exp_name] \
                --k [numbers_of_distributions] \
                --img_file [training_image_path]

Notes of training process:

• Our code supports

Inference

python test.py --name [exp_name] \
               --k [numbers_of_distributions] \
               --which_iter [which_iterations_to_load] \
               --img_file [training_image_path] \
               --sample_num [number_of_diverse_results_to_sample]

Notes of inference:

• --sample_num: How many completion results do you want?

Citation

If you find our work useful for your research, please consider citing the following papers :)

@inproceedings{xia2022pluralistic,
  title={Pluralistic Image Completion with Probabilistic Mixture-of-Experts},
  author={Xia, Xiaobo and Yang, Wenhao and Ren, Jie and Li, Yewen and Zhan, Yibing and Han, Bo and Liu, Tongliang},
  booktitle={NeurIPS},
  year={2022}
}

Acknowledgments

The authors would give special thanks to Mingrui Zhu (Xidian University), Zihan Ding (Princeton University), and Chenlai Qian (Southeast University) for helpful discussions and comments.

Contact

This repo is currently maintained by Xiaobo Xia and Wenhao Yang, which is only for academic research use. If you have any problems about the implementation of our code, feel free to contact [email protected]. Discussions and questions about our paper are welcome via [email protected].