ARCHE: Autoregressive Residual Compression with Hyperprior and Excitation

A team of researchers led by Sofia Iliopoulou, Dimitris Ampeliotis, and Athanassios Skodras has introduced ARCHE (Autoregressive Residual Compression with Hyperprior and Excitation), a new AI model for image compression. The framework unifies hierarchical, spatial, and channel-based priors to capture dependencies in image data, using adaptive feature recalibration and residual refinement to enhance quality. Crucially, it achieves state-of-the-art results without relying on computationally heavy recurrent or transformer components, making it notably efficient.

On the Kodak benchmark, ARCHE reduces the BD-Rate by approximately 48% compared to the standard model by Balle et al., 30% versus a channel-wise autoregressive model, and 5% against the advanced VVC Intra codec. It maintains this performance with 95 million parameters and a fast processing time of 222 milliseconds per image. Visual evaluations confirm the model produces sharper textures and superior color fidelity, especially at lower bit rates, demonstrating that accurate entropy modeling is feasible with streamlined, convolutional architectures suitable for real-world deployment.