Research map / Architectures and efficiency / Attention and transformers
Causal Context Adjustment Loss for Learned Image Compression
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 a Causal Context Adjustment (CCA) loss that uses an auxiliary entropy model to push informative content into early autoregressive slices, combined with a CNN-based LIC model using uneven channel-wise grouping for lower latency.
Problem
Learned image compression autoregressive entropy models rely on fixed hand-crafted causal context orders, and transformer-based encoders are slow on current hardware.
Approach
An auxiliary entropy model that sees only the hyperprior is trained alongside the main model, and a loss rewards the gap between the auxiliary and main predictions, so the encoder places more information in early causal-context slices; the model is a convolutional (NAF-block) codec with unevenly sized channel slices.
Key results
- CCA loss improves BD-rate of the uneven ChARM-style model from -14.56% to -17.17% on Kodak (BPG anchor) with no extra test-time cost.
- The full model reaches -13.87% BD-rate on Kodak versus VVC with 201 ms total latency, against 255 ms for Liu et al. at -11.88%.
- The authors report more than 20% less runtime than a recent Transformer-based SOTA model at about 2% BD-rate gain.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CCA loss to the evenly grouped channel-wise model (ablation) | ChARM (even) · Kodak | BD-rate: -13.31% -> -14.72% | Helped |
| add CCA loss to the unevenly grouped channel-wise model (ablation) | ChARM (uneven) · Kodak | BD-rate: -14.56% -> -17.17% | Helped |
| add CCA loss to the evenly grouped channel-wise model (ablation) | ChARM (even) · Kodak | Inference time: 126 -> 126 | No clear effect |
| add CCA loss to the unevenly grouped channel-wise model (ablation) | ChARM (uneven) · Kodak | Inference time: 116 -> 116 | No clear effect |
| use the unevenly grouped schedule instead of the evenly grouped schedule (ablation) | ChARM · Kodak | BD-rate: -13.31% -> -14.56% | Helped |
| use the unevenly grouped schedule instead of the evenly grouped schedule (ablation) | ChARM · Kodak | Inference time: 126 -> 116 | Helped |
| use the unevenly grouped schedule with CCA loss instead of the evenly grouped schedule with CCA loss (ablation) | ChARM + CCA loss · Kodak | BD-rate: -14.72% -> -17.17% | Helped |
| use the proposed CNN-based LIC model instead of Liu et al. | LIC model · Kodak | BD-rate: -11.88% -> -13.87% | Helped |
| use the proposed CNN-based LIC model instead of Liu et al. | LIC model · Kodak | Total inference latency: 255 -> 201 | Helped |
| select the uneven channel schedule with k=1.7 (ablation) | Proposed LIC model · Rate-distortion evaluation | Rate-distortion performance | Helped |
Surprising findings
- The CCA loss benefit is larger for the uneven grouping than for the even grouping (about 2.6 vs 1.4 BD-rate points).
Limitations
- The information redistribution effect of CCA loss and uneven grouping is not analyzed in depth, as the authors note.
- Ablations use a tiny model (halved channels and NAF-block count), so their magnitude may not carry over to the full model.
Experimental setup
Datasets: Open Images
Benchmarks: Kodak Tecnick CLIC Professional Validation
Metrics: PSNR MS-SSIM BD-rate bpp inference latency FLOPs parameters
Baselines: VVC BPG WebP JPEG Ballé2018 Zou et al. Zhu et al. Liu et al.
Builds on: hyperprior VAE compression (Ballé et al.) ChARM channel-wise autoregressive model NAFNet latent residual prediction unevenly grouped context model (He et al.)
Closest papers by meaning
- Diversify, Contextualize, and Adapt: Efficient Entropy Modeling for Neural Image Codec (NeurIPS 2024 · similarity 0.86)
- Few-Shot Domain Adaptation for Learned Image Compression (AAAI 2025 · similarity 0.86)
- Linear Attention Modeling for Learned Image Compression (CVPR 2025 · similarity 0.86)
- Balanced Rate-Distortion Optimization in Learned Image Compression (CVPR 2025 · similarity 0.85)
- GLIC: General Format Learned Image Compression (AAAI 2025 · similarity 0.85)
- Frequency-Biased Synergistic Design for Image Compression and Compensation (CVPR 2025 · similarity 0.85)
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.