Research map / Optimization / Optimizers
Balanced Rate-Distortion Optimization in 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.
Reformulates rate-distortion optimization in learned image compression as a multi-objective problem and proposes two balanced gradient-weighting methods (trajectory gradient descent for training from scratch, and an analytic QP for fine-tuning), giving about 2% BD-Rate savings.
Problem
Standard R-D loss summation in learned image compression leads to imbalanced gradient updates, where either the rate or distortion objective dominates, hurting compression efficiency.
Approach
Treat rate and distortion as two objectives in an MGDA-style multi-objective framework, adaptively weighting their log-loss gradients to maximize the minimum relative improvement speed; one solution refines weights iteratively along the training trajectory, the other solves a 2-variable equality-constrained QP analytically.
Key results
- Solution 1 on ELIC gives BD-Rate reductions of -2.68%, -2.63%, and -3.00% on Kodak, Tecnick, and CLIC2022 versus standard training.
- Solution 2 on ELIC (fine-tuning) gives -1.81%, -2.03%, and -2.39% BD-Rate reductions.
- Solution 1 increases epoch training time by about 20% and Solution 2 by 47-50%, with inference unchanged.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Balanced R-D optimization, Solution 1 (coarse-to-fine gradient descent) vs standard R-D training | ELIC, M&S Hyperprior, TCM-S · Kodak, Tecnick, CLIC2022 | BD-Rate vs standard: -1.95% to -3.00% BD-Rate | Helped |
| Balanced R-D optimization, Solution 2 (QP analytical weights) for fine-tuning vs standard R-D | ELIC, M&S Hyperprior, TCM-S · Kodak, Tecnick, CLIC2022 | BD-Rate vs standard: -1.55% to -2.39% BD-Rate | Helped |
| Remove weight renormalization (ct scaling) from the proposed solutions (ablation) | Solution 1 and Solution 2 on M&S Hyperprior · Kodak, lambda=0.013 | R-D curve convergence | Helped |
| Weight decay gamma=0 (removed) in Solution 1 softmax logit update (ablation) | Solution 1 on M&S Hyperprior · Kodak, lambda=0.013 | convergence | Hurt |
| Weight decay gamma=0.001 (vs other tested values) (ablation) | Solution 1 on M&S Hyperprior · Kodak, lambda=0.013 | R-D performance | Helped |
| Use Solution 1 for fine-tuning instead of Solution 2 (ablation) | M&S Hyperprior fine-tuning · Kodak, lambda=0.013 | R-D performance | Hurt |
| Use Solution 1 for training from scratch instead of Solution 2 (ablation) | M&S Hyperprior from scratch · Kodak, lambda=0.013 | R-D performance | No clear effect |
| Apply Solution 2 after Solution 1 (cascaded) on M&S Hyperprior (ablation) | Solution 1 models · Kodak, lambda=0.013 | R-D performance | No clear effect |
| Balanced R-D optimization (Solution 1) training overhead | M&S Hyperprior, ELIC, TCM-S · epoch training time | training time increase: +20% to +23% (Sol. 1); +47% to +50% (Sol. 2) | Hurt |
Surprising findings
- Using Solution 1 to fine-tune pre-trained models performs worse than Solution 2 for fine-tuning, while Solution 2 from scratch matches Solution 1 from scratch.
Limitations
- Adds an extra subproblem that moderately increases training complexity.
- Softmax projection of weights used instead of exact simplex projection.
Experimental setup
Datasets: COCO 2017
Benchmarks: Kodak Tecnick CLIC 2022
Metrics: BD-Rate bpp-PSNR R-D curves training time overhead
Baselines: Standard R-D optimization M&S Hyperprior ELIC TCM-S
Builds on: MGDA M&S Hyperprior ELIC TCM CompressAI
Closest papers by meaning
- Learned Image Compression via Sparse Attention and Adaptive Frequency (CVPR 2026 · similarity 0.88)
- Noise-Regularized Training for Learned Image Compression (NeurIPS 2026 · similarity 0.88)
- Few-Shot Domain Adaptation for Learned Image Compression (AAAI 2025 · similarity 0.88)
- On Disentangled Training for Nonlinear Transform in Learned Image Compression (ICLR 2025 · similarity 0.87)
- Causal Context Adjustment Loss for Learned Image Compression (NeurIPS 2024 · similarity 0.85)
- GLIC: General Format Learned Image Compression (AAAI 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.