Efficient Generative Image Compression with Quantized Latents Refinement

Key points

• We observe that, for lightweight GIC methods at low bitrates, coarse deterministic quantization can collapse nearby encoder latents corresponding to different fine details to the same point, making these details difficult to recover.1
• Motivated by this, we propose GR-GIC, an efficient GIC that delivers competitive reconstruction performance with fast inference.1
• For better performance, we introduce Guided Refinement (GR), which combines latent refinement guided by entropy context with adaptive perturbations for training regularization, jointly improving reconstruction at low bitrates.1
• For fast inference, we introduce reorganized projected residual vector quantization (P-RVQ), which equivalently reorganizes residual updates into low-dimensional lookups to expose greater parallelism and reduce sequential computation.1

Sources (1)

• [1]Efficient Generative Image Compression with Quantized Latents Refinement
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 21, 01:03 PM

  Generative image compression (GIC) delivers strong perceptual quality at low bitrates.

  We observe that, for lightweight GIC methods at low bitrates, coarse deterministic quantization can collapse nearby encoder latents corresponding to different fine details to the same point, making these details difficult to recover.

Extractive summary: sentences quoted from the sources.