Re2IC: Realism-Enhanced Region-Based Implicit Codec with Wavelet–Wasserstein Distortion

Haotian Wu, Gen Li, Di You, Pier Luigi Dragotti, Deniz Gunduz

Keywords: Implicit codec, perceputal compression, image compression, overfitted codec, low-complexity compression

Abstract: Recent advances in perceptual image compression improve quality through powerful probabilistic models but often incur prohibitive decoding costs. This paper proposes a low-complexity alternative called **Re**alism-enhanced **Re**gion-based **I**mplicit **C**odec, Re$^2$IC, that models visual perception region-by-region with saliency-guided partitioning. To further enhance realism, we introduce wavelet–Wasserstein distortion (WA-WD), which balances fidelity and perception by explicitly controlling frequency characteristics, while also yielding a tighter and more reliable diagonal-Gaussian approximation than standard WD. Together, these components enable fine-grained spatial–spectral optimization, allowing Re$^2$IC to achieve superior rate–perception (RP) trade-offs and outperform generative codecs like HiFiC while using less than 1\% of their decoding cost. Re$^2$IC achieves faster encoding convergence, low decoding, and high-quality reconstructions that preserve both fine textures and natural details. Experiments and user studies show that Re$^2$IC sets a new state of the art in RP performance among overfitted codecs. Beyond compression, WA-WD also serves as a standalone, tunable metric, that aligns more closely with human preference, achieving the highest Pearson (94.6\%) and Spearman (92.3\%) correlations with Elo scores, and demonstrating leading performance across multiple IQA datasets.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 12407