High-Fidelity Generative Image Compression Using Conditional Decoder

Haeyoon Yang, Nam Ik Cho

Abstract: This paper describes the method submitted by team ISPL_IC for the Challenge on Learned Image Compression 2025 (CLIC 2025). We propose a generative image compression approach that combines a conditional discriminator and high-frequency-aware objective to improve both perceptual quality and fidelity. The method is implemented using an ELIC generator and a conditional discriminator, which are trained with rate-distortion, adversarial, high-frequency, and perceptual loss terms. The architecture, training procedure, and datasets are illustrated to facilitate reproducibility.

Team Name: ISPL_IC

Submission Number: 4