Research map / Architectures and efficiency / CNN design
Joint Global and Local Hierarchical Priors for Learned Image Compression
Jun-Hyuk Kim, Byeongho Heo, Jong-Seok Lee
Check 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 |
|---|---|---|---|
| Replace the context and hyperprior entropy model with joint global and local hyperpriors | Learned image compression · Kodak and Tecnick | Rate savings relative to BPG | Improved |
| Use Informer instead of Context+Hyperprior+Global Reference | Learned image compression · Kodak | Rate savings relative to BPG: 1.74% up to 2.60 % | Improved |
| Use Informer instead of Context+Hyperprior+Global Reference | Entropy model · 4096×2304 | GFLOPs: 335.13 -> 329.17 | Improved |
| Add the global hyperprior (ablation) | Context-only entropy model (Ours w/o Global & Local) · Kodak, λ=0.0067 and λ=0.0483 | Rate-distortion performance | Improved |
| Add the local hyperprior alongside the global hyperprior (ablation) | Informer · Kodak, λ=0.0067 and λ=0.0483 | Rate-distortion performance | Improved |
| Combine Context+Hyperprior with the global hyperprior rather than decompose it into global and local hyperpriors (ablation) | Context+Hyperprior+Global · Higher bpp region | Rate-distortion performance | Hurt |
| Use the global hyperprior to update the context prior rather than the local hyperprior (ablation) | Informer parameter model · Kodak, λ=0.0067 and λ=0.0483 | Rate-distortion performance | Improved |
| Use four rather than eight global tokens (ablation) | Informer · Lower bpp region | Rate-distortion performance | Improved |
| Use eight rather than four global tokens (ablation) | Informer · Higher bpp region | Rate-distortion performance | Improved |
| Introduce global vision with masked self-attention in the context model instead of joint global and local hyperpriors (ablation) | Learned image compression · Kodak, λ=0.0067 | Rate (bpp): 0.2763 -> 0.2842 | Hurt |
| Introduce global vision with masked self-attention in the context model instead of joint global and local hyperpriors (ablation) | Learned image compression · Kodak, λ=0.0483 | Rate (bpp): 0.8514 -> 0.8718 | Hurt |
Closest papers by meaning
- Entroformer: A Transformer-based Entropy Model for Learned Image Compression (ICLR 2022 · similarity 0.91)
- Learned Image Compression with Dictionary-based Entropy Model (CVPR 2025 · similarity 0.89)
- GLIC: General Format Learned Image Compression (AAAI 2025 · similarity 0.87)
- Robustly overfitting latents for flexible neural image compression (NeurIPS 2024 · similarity 0.87)
- Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis (CVPR 2024 · similarity 0.86)
- Linear Attention Modeling for Learned Image Compression (CVPR 2025 · 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.
