Large Scale Image Completion via Co-Modulated Generative Adversarial Networks, ICLR 2021 (Spotlight)

Demo | Paper

[NOTICE] Our web demo will be closed recently. Enjoy the last days!

[NEW!] Time to play with our interactive web demo!

Numerous task-specific variants of conditional generative adversarial networks have been developed for image completion. Yet, a serious limitation remains that all existing algorithms tend to fail when handling large-scale missing regions. To overcome this challenge, we propose a generic new approach that bridges the gap between image-conditional and recent modulated unconditional generative architectures via co-modulation of both conditional and stochastic style representations. Also, due to the lack of good quantitative metrics for image completion, we propose the new Paired/Unpaired Inception Discriminative Score (P-IDS/U-IDS), which robustly measures the perceptual fidelity of inpainted images compared to real images via linear separability in a feature space. Experiments demonstrate superior performance in terms of both quality and diversity over state-of-the-art methods in free-form image completion and easy generalization to image-to-image translation.

Large Scale Image Completion via Co-Modulated Generative Adversarial Networks
Shengyu Zhao, Jonathan Cui, Yilun Sheng, Yue Dong, Xiao Liang, Eric I Chang, Yan Xu
Tsinghua University and Microsoft Research
arXiv | OpenReview

Overview

This repo is implemented upon and has the same dependencies as the official StyleGAN2 repo. We also provide a Dockerfile for Docker users. This repo currently supports:

• Large scale image completion experiments on FFHQ and Places2
• Image-to-image translation experiments on Edges2Shoes and Edges2Handbags
• Image-to-image translation experiments on COCO-Stuff
• Evaluation code of Paired/Unpaired Inception Discriminative Score (P-IDS/U-IDS)

Datasets

• FFHQ dataset (in TFRecords format) can be downloaded following the StyleGAN2 repo.
• Places2 dataset can be downloaded in this website (Places365-Challenge 2016 high-resolution images, training set and validation set). The raw images should be converted into TFRecords using dataset_tools/create_from_images.py with --shuffle --compressed.
• Edges2Shoes and Edges2Handbags datasets can be downloaded following the pix2pix repo. The raw images should be converted into TFRecords using dataset_tools/create_from_images.py with --shuffle --pix2pix.
• To prepare a custom dataset, please use dataset_tools/create_from_images.py, which will automatically center crop and resize your images to the specified resolution. You only need to specify --val-image-dir for testing purpose.

Training

The following script is for training on FFHQ. It will split 10k images for validation. We recommend using 8 NVIDIA Tesla V100 GPUs for training. Training at 512x512 resolution takes about 1 week.

python run_training.py --data-dir=DATA_DIR --dataset=DATASET --metrics=ids10k --mirror-augment --num-gpus=8

The following script is for training on Places2, which has a validation set of 36500 images:

python run_training.py --data-dir=DATA_DIR --dataset=DATASET --metrics=ids36k5 --total-kimg 50000 --num-gpus=8

The following script is for training on Edges2Handbags (and similarly for Edges2Shoes):

python run_training.py --data-dir=DATA_DIR --dataset=DATASET --metrics=fid200-rt-handbags --mirror-augment --num-gpus=8

Pre-Trained Models

Our pre-trained models are available on Google Drive:

Use the following script to run the interactive demo locally:

python run_demo.py -d DATA_DIR/DATASET -c CHECKPOINT_FILE(S)

or the following command as a minimal example of usage:

python run_generator.py -c CHECKPOINT_FILE -i imgs/example_image.jpg -m imgs/example_mask.jpg -o imgs/example_output.jpg

Evaluation

The following script is for evaluation:

python run_metrics.py --data-dir=DATA_DIR --dataset=DATASET --network=CHECKPOINT_FILE(S) --metrics=METRIC(S) --num-gpus=1

Commonly used metrics are ids10k and ids36k5 (for FFHQ and Places2 respectively), which will compute P-IDS and U-IDS together with FID. By default, masks are generated randomly for evaluation, or you may append the metric name with -h0 ([0.0, 0.2]) to -h4 ([0.8, 1.0]) to specify the range of masked ratio.

Citation

If you find this code helpful, please cite our paper:

@inproceedings{zhao2021comodgan,
  title={Large Scale Image Completion via Co-Modulated Generative Adversarial Networks},
  author={Zhao, Shengyu and Cui, Jonathan and Sheng, Yilun and Dong, Yue and Liang, Xiao and Chang, Eric I and Xu, Yan},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2021}
}