PerQ: Inverse Generative Modeling for Neural Image Compression via Quantization Error Compensation

Key points

• Neural image codecs face a fundamental trade-off between pixel fidelity and perceptual quality.1
• Distortion-optimized codecs preserve pixel-wise fidelity but produce blurred reconstructions at low bitrates, whereas generative codecs leverage learned priors to improve perceptual quality but tend to exhibit saturating PSNR at higher bitrates and require separate models for different bitrate targets.1
• This motivates performing generative reconstruction within the codec-induced quantization cell, rather than sampling from an unconstrained image prior.1
• Experiments on Kodak and CLIC2020 show that PerQ achieves perceptual performance comparable to generative codecs such as MS-ILLM in LPIPS and FID, while maintaining competitive PSNR relative to distortion-optimized codecs such as ELIC.1

Sources (1)

• [1]PerQ: Inverse Generative Modeling for Neural Image Compression via Quantization Error Compensation
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 10:57 PM

  We propose PerQ, which reframes perceptual reconstruction as inverse generative modeling with bounded support.

  Neural image codecs face a fundamental trade-off between pixel fidelity and perceptual quality.

Extractive summary: sentences quoted from the sources.