This is the official PyTorch implementation of
"Controlling Rate, Distortion, and Realism: Towards a Single Comprehensive Neural Image Compression Model (WACV2024, Oral)".
Shoma Iwai, Tomo Miyazaki, and Shinichiro Omachi

We have proposed a single image compression model that can control bitrate, distortion, and realism. As shown below, our method can cover a wide range of rate-distortion-realism points within a single model. Please refer to our paper for more details.

curl -sSL https://install.python-poetry.org | python3 -

We used poetry==1.6.1.

https://github.com/iwa-shi/CRDR.git

poetry install

PyTorch 1.12.1
CUDA 11.3
CompressAI 1.2.4

Other PyTorch/CUDA versions might work, but these are the environments that we tested.

You can download it from GDrive manually or use the following command:

curl "https://drive.usercontent.google.com/download?id=1H6T9-k0RX5SXk0VljHNiXUGXZrZl2seb&confirm=xxx" -o crdr.pth.tar

You can run a quick test on the images in ./demo_images (three images from Kodak dataset) using the following command:

poetry run python scripts/compress.py --config_path ./config/crdr.yaml --model_path ./crdr.pth.tar --img_dir ./demo_images --save_dir ./demo_results/crdr_q000_b384_kodak -q 0.00 -b 3.84 --decompress -d cuda

Binary compressed files (kodimXXX.bin), reconstructions (kodimXXX.png), _bitrates.csv, and _avg_bitrate.json will be stored in ./demo_results/crdr_q000_b384_kodak.

The average bitrate should be avg_bpp: 0.0641. Please verify that the reconstructions look correct.

You can run the pre-trained model on your dataset, such as the CLIC2020 test dataset and Kodak dataset.

poetry run python scripts/compress.py --config_path ./config/crdr.yaml --model_path ./crdr.pth.tar --img_dir PATH/TO/DATASET --save_dir ./results/crdr_qXXX_bXXX_DATASET -q QUALITY -b BETA --decompress -d cuda

• --quality adjusts the bitrate. Float value from 0.0 to 4.0 in our pre-trained model. 0.0: Low bitrate, 4.0: High bitrate
• --beta adjusts realism. Float value from 0.0 to 5.12. 0.0: Low distortion, 5.12: High realism

In Fig.5 in our paper, --quality: {0.0, 0.25, 0.5, ..., 3.5, 3.75, 4.0}, --beta: {0.0, 3.84} are used (17*2=34 points in total).

For example,

poetry run python scripts/compress.py --config_path ./config/crdr.yaml --model_path ./crdr.pth.tar --img_dir ./datasets/CLIC/test --save_dir ./results/crdr_q150_b384_CLIC -q 1.5 -b 3.84 --decompress -d cuda

You can calculate metrics (PSNR, FID, LPIPS, DISTS) of the reconstructions by running the following script:

poetry run python scripts/calc_metrics.py --real_dir PATH/TO/DATASET --fake_dir PATH/TO/RECONSTRUCTIONS -d cuda

Results will be stored in {fake_dir}/_metrics.json.

For example,

poetry run python scripts/calc_metrics.py --real_dir ./datasets/CLIC/test --fake_dir ./results/crdr_q150_b384_CLIC -d cuda

You can find results on CLIC2020, Kodak, and DIV2K at rd_results.

See ./docs/training.md.

• Release repository
• Pre-trained model
• Test code and instructions
• Training code and instructions

If you find this code useful for your research, please consider citing our paper:

@INPROCEEDINGS{iwai2024crdr,
  author={Iwai, Shoma and Miyazaki, Tomo and Omachi, Shinichiro},
  booktitle={2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)}, 
  title={Controlling Rate, Distortion, and Realism: Towards a Single Comprehensive Neural Image Compression Model}, 
  year={2024},
  volume={},
  number={},
  pages={2888-2897},
  doi={10.1109/WACV57701.2024.00288}}

We thank the authors of the following repositories:

• CompressAI: For entropy_model.
• ELIC (unofficial implementation by JiangWeibeta)
• HiFiC
• HiFiC (unofficial PyTorch Implementation by Justin-Tan)
• tensorflow-compression
• MMEngine: For config and registry.

If you have any questions or encounterd any issues, please feel free to open issue.