Research map / Architectures and efficiency / RNNs and normalization
Diversify, Contextualize, and Adapt: Efficient Entropy Modeling for Neural Image Codec
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 DCA, an entropy model for neural image codecs that diversifies forward-adaptation hyper latents into local, regional, and global types and contextualizes them step-adaptively on top of quadtree-based backward adaptation, reporting 3.73% BD-rate gain over the prior state of the art on Kodak.
Problem
Efficient backward-adaptation entropy models (quadtree partition) are limited because forward adaptation uses only a single regional hyper latent, providing insufficient context especially in the first modeling step, which hurts rate-distortion performance.
Approach
Diversify the hyper latent representations into local, regional, and global types for forward adaptation, then contextualize each step with these in a step-adaptive manner using the order regional→global→local, so the first step gets richer context without extra bitrate.
Key results
- DCA achieves 3.73% BD-rate gain over the quadtree-based state-of-the-art on Kodak (vs VTM-12.1 baseline).
- DCA achieves 11.96% average rate savings over VTM-12.1 on Kodak, versus 8.55% for Lie et al. 2023 and 4.86% for Minnen & Singh 2020.
- DCA decoding time is 82.05 ms with 37.89M parameters, versus 67.05 ms and 32.64M for the quadtree baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DCA instead of the quadtree-based entropy model | ELIC-sm image codec · Kodak | BD-rate: 3.73% BD-rate gain | Helped |
| Use DCA instead of the competing methods | ELIC-sm image codec · Kodak | BD-rate: $-$3.73 vs 0.00 | Helped |
| Add global and local contexts to regional forward context (ablation) | DCA · Two bit-rate regions | Rate-distortion performance | Helped |
| Use a larger regional context instead of diversified contexts (ablation) | DCA · Two bit-rate regions | Rate-distortion performance | No clear effect |
| Replace diversified regional, global, and local contexts with global and local contexts (ablation) | DCA · Two bit-rate regions | Compression efficiency | Hurt |
| Add global context to quadtree backward adaptation (ablation) | DCA · Two bit-rate regions | Rate-distortion performance | No clear effect |
| Use step-adaptive regional and global contexts instead of combining them in advance (ablation) | DCA · Two bit-rate regions | Rate-distortion performance | Helped |
| Prioritize regional context in the modeling order (ablation) | DCA contextualization · Two bit-rate regions | Rate-distortion performance | Helped |
| Use the opposite order, local then global then regional, instead of the regional-first order (ablation) | DCA contextualization · Lower bit-rate region | Rate-distortion performance | Hurt |
| Use CNN for forward adaptation and attention for backward adaptation (ablation) | DCA architecture · Two bit-rate regions | Rate-distortion performance | Helped |
Surprising findings
- Using additional global information for backward adaptation gave no distinguishable advantage over the local+regional quadtree design.
- Simply enlarging the regional context for forward adaptation did not improve performance.
Limitations
- Architectures are inherited from existing works rather than designed for diverse contexts.
- Diversification considered only along spatial range (local/regional/global).
Experimental setup
Datasets: OpenImages
Benchmarks: Kodak Tecnick
Metrics: BD-rate PSNR bpp decoding time number of parameters
Baselines: Quadtree (Lie et al. 2023) Minnen & Singh 2020 ChARM LIC-TCM MLIC++ VTM-12.1
Builds on: quadtree partition-based backward adaptation ELIC Informer hyperprior joint backward and forward adaptation
Closest papers by meaning
- Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis (CVPR 2024 · similarity 0.88)
- Robustly overfitting latents for flexible neural image compression (NeurIPS 2024 · similarity 0.87)
- HyperVQ: Enabling Hyperprior Entropy Modeling for VQ-Based Generative Image Compression (NeurIPS 2026 · similarity 0.87)
- Causal Context Adjustment Loss for Learned Image Compression (NeurIPS 2024 · similarity 0.86)
- Controllable Distortion-Perception Tradeoff Through Latent Diffusion for Neural Image Compression (AAAI 2025 · similarity 0.86)
- Learned Image Compression with Dictionary-based Entropy Model (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.