Research map / Image generation / Image restoration
An Information-Theoretic Regularizer for Lossy Neural Image Compression
Yingwen Zhang, Meng Wang, Xihua Sheng, Peilin Chen, Junru Li, Li Zhang, Shiqi Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add negative conditional source entropy regularizer to the rate-distortion loss | MLIC++ · Natural images average | BD-rate: -3.03% | Improved |
| add the regularizer to classic hyperprior, autoregressive and attention models | hyperprior, autoregressive, attention · Natural images average | BD-rate: -0.88%, -1.06%, -0.92% | Improved |
| apply the regularizer to models evaluated on unseen domains | all five compression models · Out-of-domain (pixel, screen, game, pathology) | BD-rate: -0.43% to -3.44% | Improved |
| apply the regularizer on the attention model for pathology versus natural images | attention · Pathology | BD-rate: -0.92% -> -2.38% | Improved |
| train longer to 4x10^6 steps with the regularizer (ablation) | hyperprior · Kodak natural images | BD-rate: -1.24% at 2.5M -> around 0.7% gain at 4M | Mixed |
| set regularization factor alpha too large or too small | hyperprior · Kodak natural images | BD-rate | Hurt |
| add an H(U|X_hat) estimator to the regularizer | regularizer · Kodak natural images | BD-rate | Hurt |
| use a stronger autoregressive source entropy model for a hyperprior compression model (ablation) | hyperprior · Kodak natural images | BD-rate: average BD-Rate remains positive | Hurt |
| use a weaker hyperprior source entropy model for an autoregressive compression model (ablation) | autoregressive · Kodak natural images | BD-rate: -0.20% at 1.5M steps | Mixed |
| train an additional source entropy model during training | MLIC++ · Training time | relative training time: 100% -> 129.8% | Hurt |
Closest papers by meaning
- Improving Inference for Neural Image Compression (NeurIPS 2020 · similarity 0.90)
- Robustly overfitting latents for flexible neural image compression (NeurIPS 2024 · similarity 0.89)
- Joint Global and Local Hierarchical Priors for Learned Image Compression (CVPR 2022 · similarity 0.89)
- Noise-Regularized Training for Learned Image Compression (NeurIPS 2026 · similarity 0.89)
- Soft then Hard: Rethinking the Quantization in Neural Image Compression (ICML 2021 · similarity 0.89)
- Learned Image Compression with Dictionary-based Entropy Model (CVPR 2025 · similarity 0.88)
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.
