Research map / Probabilistic and scientific modeling / Variational and generative sampling
Synonymous Variational Inference for Perceptual Image Compression
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes synonymous variational inference (SVI) and a progressive synonymous image compression (SIC) codec that reframes perceptual image compression as a synonymous rate-distortion-perception tradeoff, evaluated against HiFiC and MS-ILLM.
Problem
Perceptual image compression lacks a unified theoretical foundation; existing rate-distortion-perception methods use inconsistent divergence/perception measures (KL, adversarial, LPIPS, DISTS) without a principled explanation of why distribution divergence appears in the objective.
Approach
Model perceptual similarity as a synonymy criterion building an ideal synonymous set, and derive via variational inference on a partial semantic KL divergence that the perceptual compression objective is a synonymous rate-distortion-perception tradeoff; implement a progressive codec that encodes only the synonymous part of the latent and samples details to generate a synset of reconstructions.
Key results
- A single progressive SIC codec with 16 synonymous levels achieves comparable rate-distortion-perception performance and surpasses the No-GAN MS-ILLM on DISTS over a large coding-rate range.
- HiFiC and MS-ILLM (GAN-based) retain a clear advantage on DISTS and especially FID; the paper reports a notable gap on FID.
- Fine-tuning with an adversarial loss improves DISTS at higher bitrates but reduces PSNR, and FID gains are limited.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use progressive SIC instead of No-GAN MS-ILLM | Progressive SIC (M=1) · CLIC2020, DIV2K, Kodak | DISTS | Helped |
| Increase reconstructed samples during training from M=1 to M=5 (ablation) | Progressive SIC · CLIC2020, DIV2K, Kodak | DISTS | Helped |
| Fine-tune the synthesis transform with GAN adversarial loss (ablation) | Progressive SIC (M=1 and M=5) · CLIC2020, DIV2K, Kodak | DISTS | Helped |
| Fine-tune the synthesis transform with GAN adversarial loss (ablation) | Progressive SIC (M=1 and M=5) · CLIC2020, DIV2K, Kodak | FID | Helped |
| Fine-tune the synthesis transform with GAN adversarial loss (ablation) | Progressive SIC (M=1 and M=5) · CLIC2020, DIV2K, Kodak | LPIPS | No clear effect |
| Fine-tune the synthesis transform with GAN adversarial loss (ablation) | Progressive SIC (M=1 and M=5) · CLIC2020, DIV2K, Kodak | PSNR | Hurt |
Surprising findings
- Adding adversarial fine-tuning improves DISTS and FID at higher bitrates but the FID improvement remains limited, while PSNR degrades.
Limitations
- Uses LPIPS in place of the KL divergence term, which is the main limitation identified by the authors.
- Gap to GAN-based HiFiC and MS-ILLM remains, especially on FID.
- Equal channel slicing for synonymous level partitioning is suboptimal at low bitrates.
Experimental setup
Datasets: OpenImages V6 CLIC2020 DIV2K Kodak
Benchmarks: CLIC2020 DIV2K Kodak
Metrics: PSNR LPIPS DISTS FID
Baselines: BPG VTM HiFiC MS-ILLM MS-ILLM No-GAN
Builds on: variational autoencoder rate-distortion-perception tradeoff joint autoregressive and hierarchical prior semantic information theory (synonymity)
Closest papers by meaning
- Towards Unified Human Perception and Machine Understanding: Token Flow Guided Compression Framework (CVPR 2026 · similarity 0.86)
- Generative Latent Coding for Ultra-Low Bitrate Image Compression (CVPR 2024 · similarity 0.85)
- Ultra Lowrate Image Compression with Semantic Residual Coding and Compression-aware Diffusion (ICML 2025 · similarity 0.85)
- Neural Image Compression with Text-guided Encoding for both Pixel-level and Perceptual Fidelity (ICML 2024 · similarity 0.85)
- Large Language Models for Lossless Image Compression: Next-Pixel Prediction in Language Space is All You Need (NeurIPS 2025 · similarity 0.84)
- Robustness-Guided Image Synthesis for Data-Free Quantization (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.