Research map / Label-efficient and robust learning / Test-time adaptation
Test-time Adaptation for Image Compression with Distribution Regularization
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 distribution-regularized hybrid latent refinement for cross-domain test-time adaptation image compression, using dropout-based Bayesian approximation of the hyperprior posterior to reduce rate cost.
Problem
Hybrid latent refinement (HLR), effective for in-domain test-time adaptation of learned image compression, increases rate cost under cross-domain shifts (e.g., natural to screen content images), due to mismatch between refined Gaussian conditional and hyperprior distributions.
Approach
Adds a distribution regularization term -log p(z|y) to the HLR objective, with the posterior p(z|y) estimated via dropout variational inference (Monte Carlo dropout on the hyperprior analysis transform) so that no new network is needed; this encourages a better joint probability approximation of the marginal and lowers rate in cross-domain settings.
Key results
- On cross-domain datasets with the Hyperprior backbone, the proposed method reaches an average BD-rate of -20.78% versus -14.87% for HLR and -15.41% for BLR.
- With the AR-CM backbone, the average BD-rate is -27.00% for the proposed method versus -21.36% for HLR and -21.41% for BLR.
- Plugged into the WACNN-based TTA-IC pipeline, the stage-1 average BD-rate improves from -10.02% (HLR) to -12.01%, and the full two-stage average improves from -14.89% to -14.89 -> -16.95% with Adaptor1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add distribution regularization to HLR | AR-CM · Cross-domain datasets (SIQAD, SCID, CCT, Pixel, DomainNet) | BD-rate savings (%): -21.41 -> -27.00 | Helped |
| Add distribution regularization to HLR | Hyperprior · Cross-domain datasets (SIQAD, SCID, CCT, Pixel, DomainNet) | BD-rate savings (%): -15.41 -> -20.78 | Helped |
| Replace vanilla HLR with distribution-regularized HLR | SOTA TTA-IC Stage 1 (WACNN) · DomainNet | BD-rate savings (%): -2.94 -> -5.46 | Helped |
| Increase the number of dropout variational inference layers (ablation) | Distribution regularization · SIQAD | BD-rate savings (%) | Helped |
| Set regularization coefficient to β=10 (ablation) | Distribution regularization · SIQAD | BD-rate savings (%) | Hurt |
| Remove Bayesian approximation and use deterministic distribution regularization (σ = 1, no dropout layer) (ablation) | Distribution regularization · SIQAD | BD-rate savings (%) | Hurt |
| Use bit-back coding instead of the proposed distribution regularization (ablation) | Cross-domain image compression · Cross-domain datasets | BD-rate savings (%) | No clear effect |
Surprising findings
- Vanilla HLR gives better reconstruction than BLR in cross-domain tests but consumes more bits, and performs below BLR on several cross-domain tasks (e.g., SCID, CCT, DomainNet).
- Bit-back coding yields only negligible gains on cross-domain image compression.
Limitations
- Gains are smaller on some backbones and tasks for individual datasets (e.g., Hyperprior on Pixel is weaker than AR-CM), and the paper reports only one dropout probability (0.5) and T=20.
- Adds 2000 Adam iterations per image at test time, so the adaptation cost remains substantial.
- Medical image generalization is shown only qualitatively and briefly.
Experimental setup
Datasets: Kodak SIQAD SCID CCT Pixel (self-collected gaming images) DomainNet
Benchmarks: Kodak SIQAD SCID CCT Pixel DomainNet
Metrics: BD-rate PSNR bpp R-D curves
Baselines: BLR HLR BBC WACNN + HLR AR-CM + BLR Hyperprior + BLR Adaptor1 Adaptor2
Builds on: hybrid latent refinement basic latent refinement dropout variational inference bit-back coding CompressAI
Closest papers by meaning
- Robustly overfitting latents for flexible neural image compression (NeurIPS 2024 · similarity 0.86)
- Few-Shot Domain Adaptation for Learned Image Compression (AAAI 2025 · similarity 0.86)
- OSCAR: One-Step Diffusion Codec Across Multiple Bit-rates (NeurIPS 2025 · similarity 0.85)
- Adapt and Diffuse: Sample-adaptive Reconstruction via Latent Diffusion Models (ICML 2024 · similarity 0.85)
- Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers (CVPR 2025 · similarity 0.85)
- Backpropagation-Free Test-Time Adaptation via Probabilistic Gaussian Alignment (NeurIPS 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.