Research map / Image generation / Image restoration
Good, Cheap, and Fast: Overfitted Image Compression with Wasserstein Distortion
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.
Optimizes the overfitted C3 neural codec with a Wasserstein Distortion (VGG-feature, saliency-modulated) loss plus common randomness, achieving HiFiC-like human-rated quality at under 1% of decoder MACs.
Problem
Learned image compression with strong perceptual quality usually relies on large generative models that are too computationally expensive to decode on practical devices; the goal is high perceptual quality at low bit rate with low decoding complexity.
Approach
Replace the MSE distortion of the overfitted C3 codec with an approximate Wasserstein Distortion computed on VGG features with a saliency-derived sigma map (discretized to powers of two for efficiency), and supply pseudo-random common randomness shared by encoder and decoder so textures can be reproduced by noise shaping.
Key results
- C3/WD achieves human-rated quality and bit-rate trade-off comparable to HiFiC while requiring over two orders of magnitude fewer decoder MACs.
- WD8 predicts Elo scores of 29 method/rate combinations with Pearson correlation 0.936 and WDs (sigma_max=16) reaches 0.942, versus 0.764 for PIM-5 and 0.711 for LPIPS.
- On binary crop ratings, WDs predicts 72.87% of human choices correctly, compared with 69.96% for LPIPS and 61.27% for PSNR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use WD instead of LPIPS as the image-quality predictor | Perceptual metrics evaluated against human Elo scores · CLIC2020 professional validation | PCC with Elo scores: 0.711 -> 0.936 | Helped |
| Optimize C3 for WD instead of MSE | C3 · CLIC2020 professional validation, human preference | Elo score | Helped |
| Use saliency-derived σ instead of constant σ (ablation) | C3 optimized for WD · CLIC2020 professional validation, high bit rates | visual quality | Helped |
| Add common randomness (ablation) | C3 optimized for WD · CLIC2020 professional validation, low bit rates | visual quality | Helped |
| Use saliency-weighted MSE instead of unweighted MSE (ablation) | C3 · CLIC2020 professional validation | Elo score | Helped |
| Use WD with a saliency-derived σ-map instead of constant σ (ablation) | C3 optimized for WD · CLIC2020 image 32 crop | text legibility | Helped |
| Add common randomness to WD-optimized C3 (ablation) | C3 optimized for WD · CLIC2020 image 28 crop | texture quality | Helped |
| Optimize C3 for WD rather than MSE | C3 · CLIC2020 professional validation | decoder MACs | No clear effect |
Surprising findings
- Adding common randomness mainly helps at low bit rates while removing saliency hurts mainly at high bit rates.
- WD optimization reduces the fraction of bits spent on the highest-resolution latent array, indicating texture resampling rather than detail coding.
Limitations
- Encoding complexity is high: WD optimization increases wall time per step about 6x over MSE.
- The feature space and sigma-map choices are described as ad-hoc and need further validation.
- Human rating study uses a single dataset (CLIC2020 professional validation) and a limited set of raters and comparisons.
Experimental setup
Datasets: CLIC2020 professional validation
Benchmarks: CLIC2020 professional validation
Metrics: Elo score bits/pixel MACs/pixel percent correct Pearson correlation Spearman correlation
Baselines: C3/MSE C3/LPIPS C3/MS-SSIM VVC MLIC+ CDC HiFiC
Builds on: C3 COOL-CHIC Wasserstein Distortion LPIPS EML-Net
Closest papers by meaning
- Towards image compression with perfect realism at ultra-low bitrates (ICLR 2024 · similarity 0.89)
- High-Fidelity Generative Image Compression (NeurIPS 2020 · similarity 0.89)
- Decouple Distortion from Perception: Region Adaptive Diffusion for Extreme-low Bitrate Perception Image Compression (CVPR 2025 · similarity 0.88)
- Controllable Distortion-Perception Tradeoff Through Latent Diffusion for Neural Image Compression (AAAI 2025 · similarity 0.87)
- GLIC: General Format Learned Image Compression (AAAI 2025 · similarity 0.87)
- C3: High-Performance and Low-Complexity Neural Compression from a Single Image or Video (CVPR 2024 · similarity 0.86)
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.