Research map / Image generation / Image restoration
Learning End-to-End Lossy Image Compression: A Benchmark
Yueyu Hu, Wenhan Yang, Zhan Ma, Jiaying Liu
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 coarse-to-fine hyperprior layer (4K images) (ablation) | single-layer hyperprior · LIU4K subset 4K | BD-rate (%): -4.65% | Improved |
| add coarse-to-fine hyperprior layer (1080p) (ablation) | single-layer hyperprior · LIU4K subset 1080p | BD-rate (%): -2.65% | Improved |
| add coarse-to-fine hyperprior layer (540p) (ablation) | single-layer hyperprior · LIU4K subset 540p | BD-rate (%): -1.97% | Improved |
| aggregate hyperlatents in full-resolution IAR instead of before synthesis transform (ablation) | Mean aggregation (Fine) · Information aggregation ablation | BD-rate (%): -4.17% -> -6.35% | Improved |
| add coarse granularity hyperrepresentation to aggregation (ablation) | Mean IAR · Information aggregation ablation | BD-rate (%): -6.35% -> -7.34% | Improved |
| aggregate hyperprior representation instead of Gaussian mean (ablation) | IAR (Fine+Coarse) · Information aggregation ablation | BD-rate (%): -7.34% -> -8.38% | Improved |
| add hyperprior to factorized entropy model | ICLR18 · Kodak (full range) | BD-rate (%): 32.96% -> 4.07% | Improved |
| add hyperprior to factorized entropy model | ICLR18 · Tecnick (full range) | BD-rate (%): 36.70% -> 0.98% | Improved |
| add GDN-like hyperprior structure vs proposed coarse-to-fine model | Ours vs ICLR19 context model · Kodak (full range) | BD-rate (%): -5.40% -> -9.42% | Improved |
| proposed coarse-to-fine model vs context model method on Tecnick | Ours vs NeurIPS18 · Tecnick (full range) | BD-rate (%): -14.54% vs -14.15% | No clear effect |
| proposed model vs context model method on CLIC | Ours vs NeurIPS18 · CLIC 2019 val (full range) | BD-rate (%): -1.88% -> -12.86% | Improved |
Closest papers by meaning
- End-to-end Optimized Image Compression (ICLR 2017 · similarity 0.92)
- Learned Video Compression (ICCV 2019 · similarity 0.92)
- Compress Clean Signal from Noisy Raw Image: A Self-Supervised Approach (ICML 2024 · similarity 0.90)
- Good, Cheap, and Fast: Overfitted Image Compression with Wasserstein Distortion (CVPR 2025 · similarity 0.89)
- GLIC: General Format Learned Image Compression (AAAI 2025 · similarity 0.88)
- C3: High-Performance and Low-Complexity Neural Compression from a Single Image or Video (CVPR 2024 · 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.
