Advances in Diffusion-Based Generative Compression

TL;DR

This paper reviews recent diffusion-based generative methods for image compression, emphasizing their ability to produce realistic reconstructions at very low bit-rates and discussing theoretical and practical aspects.

Contribution

It provides a comprehensive review of diffusion-based lossy compression techniques, analyzing their encoding, decoding, and transmission strategies within a unified framework.

Findings

Diffusion models enable high-quality image reconstruction at low bit-rates.

Recent methods incorporate diffusion models into the transmission process via channel simulation.

The review highlights open challenges and theoretical insights in diffusion-based compression.

Abstract

Popularized by their strong image generation performance, diffusion and related methods for generative modeling have found widespread success in visual media applications. In particular, diffusion methods have enabled new approaches to data compression, where realistic reconstructions can be generated at extremely low bit-rates. This article provides a unifying review of recent diffusion-based methods for generative lossy compression, with a focus on image compression. These methods generally encode the source into an embedding and employ a diffusion model to iteratively refine it in the decoding procedure, such that the final reconstruction approximately follows the ground truth data distribution. The embedding can take various forms and is typically transmitted via an auxiliary entropy model, and recent methods also explore the use of diffusion models themselves for information transmission via channel simulation. We review representative approaches through the lens of rate-distortion-perception theory, highlighting the role of common randomness and connections to inverse problems, and identify open challenges.