Title: Exploiting Semantic and Pixel Representations for Ultra-Low Bitrate Image Compression

Abstract:

This paper introduces SPRDiff, a diffusion-based image compression method designed to address limitations in existing extreme compression techniques regarding the rate-distortion-perception trade-off. The proposed method incorporates both semantic and pixel representations to maintain reconstruction fidelity under ultra-low bitrate constraints.

The architecture utilizes a triple-encoder system. This system integrates high-fidelity features from pretrained distortion-oriented and semantic-oriented encoders with a frozen Variational Autoencoder (VAE) to improve latent compression and entropy modeling. Additionally, a distortion-aware reconstruction module with dual feature extraction is employed to generate coarse reconstructions that preserve primary structures and provide semantic and pixel-level conditional signals for the diffusion model.

Experimental results on benchmark datasets indicate that the method outperforms current state-of-the-art approaches at bitrates below 0.03 bits per pixel (bpp). The study reports that the method effectively preserves both perceptual quality and pixel-wise fidelity. Source code and trained models are available at https://github.com/cshw2021/SPRDiff.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC