Core Concepts
The proposed method leverages a privileged end-to-end decoder to correct the score function of a diffusion model, achieving better perceptual quality while guaranteeing distortion.
Abstract
The paper presents a diffusion-based image compression method that employs a privileged end-to-end decoder model as correction. The key highlights are:
The authors analyze the approximation error of the score function estimated by the score network when the original images are visible at the encoder side. This provides privileged information to facilitate correcting the error at the decoder side.
The authors introduce a privileged end-to-end convolutional decoder and linearly combine it with the score network via a mathematically derived factor to build an approximation of the above-mentioned error.
The linear factors used to combine the two components are transmitted with a few bits as privileged information, assisting the decoder to correct the sampling process and achieve improved visual quality.
Extensive experiments demonstrate the superiority of the proposed "CorrDiff" method in both distortion and perception compared to previous perceptual compression methods.
Stats
The paper does not contain any key metrics or important figures to support the author's key logics.
Quotes
The paper does not contain any striking quotes supporting the author's key logics.