Once-for-All: Controllable Generative Image Compression with Dynamic Granularity Adaption

We design a unified generative compression model (Control-GIC) capable of variable bitrate adaption across a broad spectrum while preserving high-perceptual fidelity reconstruction. Control-GIC allows one model for variable bitrates and once compression on an entire dataset for constrained bitrate conditions.

• The paper received an update: See https://arxiv.org/pdf/2406.00758.
• Update the pre-trained model with the codebook size of 1024.
• Improve the code, please update your local code.

git clone https://github.com/lianqi1008/Control-GIC.git
cd Control-GIC
conda create -n CGIC python=3.10.4
conda activate CGIC
pip install -r requirements.txt
(if errors occur: conda install pip=22.3.1)

The model weight can be downloaded from GoogleDrive and BaiduNetDisk.

Train

We use OpenImages-v6 dataset and randomly crop images to a uniform 256 × 256 resolution.

python main.py --config configs/config_train.yaml

Inference

We use Kodak, high-resolution DIV2K and CLIC2020 datasets to evaluate the model.

python inference.py -i input_dir -o output_dir

Fine control of bitrate

By fine-tuning the granularity ratio, the model is able to make fine adjustments to the bitrate (Evaluated on the Kodak dataset).

Constant bitrate

Once the granularity ratio is fixed, the bpp values of each image in the dataset are very close to each other, thus easily fulfilling the constrained bandwidth and bitrate conditions. For example, in one compression on the Kodak dataset, we set the ratio to (0.1, 0.8, 0.1), and obtained 24 images with bpp ranging from 0.262 to 0.277, with only slight variations.

@inproceedings{li2025onceforall,
title={Once-for-All: Controllable Generative Image Compression with Dynamic Granularity Adaptation},
author={Anqi Li and Feng Li and Yuxi Liu and Runmin Cong and Yao Zhao and Huihui Bai},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025}
}

The code references VQGAN, MoVQGAN, DQ-VAE and Huffman coding from Bhrigu Srivastava@bhrigu123. Thanks for these excellent works!