From structure to detail: A conditional diffusion framework for extremely low-bitrate image compression

《Signal Processing》 ---- Paper

---

*** If FPD-IC is helpful for you, please help star this repo. Thanks!

Testing

1. Generation Condition image

 sh ./scripts_tmp/test_S1_TCM.sh DATASET lambda   #DATASET= ["kodak", "Tecnick"] lambda = [0.00015   0.00025  0.0005 0.001]

2. Reconstructor

  (1) Modify： the "file_gt" and  "file_lq" in configs/inference_tmp/LIC_LDM.yaml to  the printed pathway after runing "1.Generation Condition image"
  (2) sh ./scripts_tmp/test_S2_LDM.sh  DATASET lambda   #DATASET= ["kodak", "Tecnick"] lambda = [0.00015   0.00025  0.0005 0.001]

Training

1. Basic Compression Model Training

 sh ./scriptsEn/train.sh 

2. DM-based Details Recovery Training

sh ./scriptsEn/train2.sh 

Note: Condition image output diretion： file_lq FPD-IC output diretion： results_tmp/DATASET_${lambda}/

Citation

Please cite us if our work is useful for your research.

@article{LI2026110480,
title = {From structure to detail: A conditional diffusion framework for extremely low-bitrate image compression},
journal = {Signal Processing},
volume = {243},
pages = {110480},
year = {2026},
issn = {0165-1684},
doi = {https://doi.org/10.1016/j.sigpro.2025.110480},
url = {https://www.sciencedirect.com/science/article/pii/S0165168425005961},
}

License

This project is released under the Apache 2.0 license.

Acknowledgement

This project is based on ControlNet, DIFFBIR and BasicSR. Thanks for their awesome work.

Contact

If you have any questions, please feel free to contact with me at 105830@xaut.edu.cn.