Image Completion Transformer (ICT)

Project Page | Paper (ArXiv) | Pre-trained Models | Supplemental Material

This repository is the official pytorch implementation of our ICCV 2021 paper, High-Fidelity Pluralistic Image Completion with Transformers.

Ziyu Wan¹, Jingbo Zhang¹, Dongdong Chen², Jing Liao^1
1City University of Hong Kong, ²Microsoft Cloud AI

🎈 Prerequisites

• Python >=3.6
• PyTorch >=1.6
• NVIDIA GPU + CUDA cuDNN

pip install -r requirements.txt

To directly inference, first download the pretrained models from Dropbox, then

cd ICT
wget -O ckpts_ICT.zip https://www.dropbox.com/s/cqjgcj0serkbdxd/ckpts_ICT.zip?dl=1
unzip ckpts_ICT.zip

Some tips:

• Masks should be binarized.
• The extensions of images and masks should be .png.
• The model is trained for 256x256 input resolution only.
• Make sure that the downsampled (32x32 or 48x48) mask could cover all the regions you want to fill. If not, dilate the mask.

🌟 Pipeline

Why transformer?

Compared with traditional CNN-based methods, transformers have better capability in understanding shape and geometry.

🚀 Training

1) Transformer

cd Transformer
python main.py --name [exp_name] --ckpt_path [save_path] \
               --data_path [training_image_path] \
               --validation_path [validation_image_path] \
               --mask_path [mask_path] \
               --BERT --batch_size 64 --train_epoch 100 \
               --nodes 1 --gpus 8 --node_rank 0 \
               --n_layer [transformer_layer #] --n_embd [embedding_dimension] \
               --n_head [head #] --ImageNet --GELU_2 \
               --image_size [input_resolution]

Notes of transformer:

• --AMP: Reduce the memory cost while training, but sometimes will lead to NAN.
• --use_ImageFolder: Enable this option while training on ImageNet
• --random_stroke: Generate the mask on-the-fly.
• Our code is also ready for training on multiple machines.

2) Guided Upsampling

cd Guided_Upsample
python train.py --model 2 --checkpoints [save_path] \
                --config_file ./config_list/config_template.yml \
                --Generator 4 --use_degradation_2

Notes of guided upsampling:

• --use_degradation_2: Bilinear downsampling. Try to match the transformer training.
• --prior_random_degree: Stochastically deviate the sequence elements by K nearest neighbour.
• Modify the provided config template according to your own training environments.
• Training the upsample part won't cost many GPUs.

⚡ Inference

We provide very covenient and neat script for inference.

python run.py --input_image [test_image_folder] \
              --input_mask [test_mask_folder] \
              --sample_num 1  --save_place [save_path] \
              --ImageNet --visualize_all

Notes of inference:

• --sample_num: How many completion results do you want?
• --visualize_all: You could save each output result via disabling this option.
• --ImageNet --FFHQ --Places2_Nature: You must enable one option to select corresponding ckpts.
• Please use absolute path.

More results

FFHQ

Places2

ImageNet

⏳ To Do

• Release training code
• Release testing code
• Release pre-trained models
• Add Google Colab

📔 Citation

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

@article{wan2021high,
  title={High-Fidelity Pluralistic Image Completion with Transformers},
  author={Wan, Ziyu and Zhang, Jingbo and Chen, Dongdong and Liao, Jing},
  journal={arXiv preprint arXiv:2103.14031},
  year={2021}
}

The real-world application of image inpainting is also ready! Try and cite our old photo restoration algorithm here.

@inproceedings{wan2020bringing,
title={Bringing Old Photos Back to Life},
author={Wan, Ziyu and Zhang, Bo and Chen, Dongdong and Zhang, Pan and Chen, Dong and Liao, Jing and Wen, Fang},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={2747--2757},
year={2020}
}

💡 Acknowledgments

This repo is built upon minGPT and Edge-Connect. We also thank the provided cluster centers from OpenAI.

📨 Contact

This repo is currently maintained by Ziyu Wan (@Raywzy) and is for academic research use only. Discussions and questions are welcome via [email protected].