Research map / Label-efficient and robust learning / Domain adaptation and generalization
Few-Shot Domain Adaptation 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 few-shot domain adaptation for learned image compression by inserting plug-and-play Conv1x1 channel re-allocation adapters and LoRA adapters into frozen pre-trained LIC models, trained on as few as 25 target-domain samples.
Problem
Pre-trained learned image compression (LIC) models degrade substantially on out-of-training-domain images (e.g., pixel art, screen content, craters, games, pathology), falling behind VVC/VTM; existing domain adaptation is restricted to GDN architectures and adds many parameters.
Approach
Observes that out-of-domain images disrupt the pre-trained channel-wise (frequency-like) energy decomposition of latents, and adapts by channel-wise re-allocation: Conv1x1 adapters (identity-initialized) in the transform and LoRA adapters in the entropy parameters network, trained with a two-stage strategy on few target samples.
Key results
- With 25 target samples, the method reduces BD-rate by 17.64% on average for ELIC and 22.16% for Cheng2020 versus pre-trained models, comparable to full-model finetune.
- Transmits only 0.37M–0.88M parameters, i.e., 0.63%–1.95% of total model parameters, versus 34.5%–94.7% for DANICE/finetune on other models.
- Outperforms DANICE in all conditions reported and brings most pre-trained LIC models to roughly VTM-22.0 level on the five out-of-domain datasets except Pixel.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply few-shot domain adaptation | Cheng2020 · Five-domain average, 25 target samples | BD-rate: -22.16% | Helped |
| Apply few-shot domain adaptation | ELIC · Five-domain average, 25 target samples | BD-rate: -17.64% | Helped |
| Apply few-shot domain adaptation | TCM · Five-domain average, 25 target samples | BD-rate: -15.24% | Helped |
| Apply few-shot domain adaptation | MLIC++ · Five-domain average, 25 target samples | BD-rate: -11.04% | Helped |
| Add LoRA-Adapters to Conv-Adapters (ablation) | ELIC · Five-domain average, 25 target samples | BD-rate: -15.40% -> -16.43% | Helped |
| Add stage 2, finetuning only Stack 4 adapters (ablation) | ELIC with Conv-Adapters and LoRA-Adapters · Five-domain average, 25 target samples | BD-rate: -16.43% -> -17.64% | Helped |
| Finetune adapters in Stacks 1-4 instead of only Stack 4 during stage 2 (ablation) | ELIC with Conv-Adapters and LoRA-Adapters · Five-domain average, 25 target samples | BD-rate: -17.64% -> -15.19% | Hurt |
| Replace Conv1×1 with GDN (ablation) | ELIC Conv-Adapters · Five-domain average, 25 target samples | BD-rate: -17.64% -> -7.37% | Hurt |
| Replace Conv1×1 with depthwise Conv3×3 (ablation) | ELIC Conv-Adapters · Five-domain average, 25 target samples | BD-rate: -17.64% -> -10.63% | Hurt |
| Replace Conv1×1 with depthwise Conv3×3 followed by Conv1×1 (ablation) | ELIC Conv-Adapters · Five-domain average, 25 target samples | BD-rate: -17.64% -> -16.39% | Hurt |
| Replace Conv1×1 with Conv3×3 (ablation) | ELIC Conv-Adapters · Five-domain average, 25 target samples | BD-rate: -17.64% -> -17.39% | Hurt |
| Deploy Conv-Adapters in Stacks 1-4 rather than Stack 1 only (ablation) | ELIC analysis and synthesis transforms · Five-domain average, 25 target samples | BD-rate: -5.46% -> -17.64% | Helped |
Surprising findings
- Adaptation with GDN-based or depthwise-only adapters performs much worse than Conv1x1 adapters despite fewer transmitted parameters.
- Pixel domain remains hard to fully compensate, staying behind VTM even after adaptation.
Limitations
- Pixel domain gap too wide to fully close to VTM
- ELIC was self-trained since official weights were unavailable
- DANICE and some baselines are restricted to GDN architectures
Experimental setup
Datasets: Kodak DIV2K CLIC 2020 Professional SCID SCI1K SIQAD LROC GamingVideoSET BRACS pixel-style art dataset
Benchmarks: Kodak Pixel Screen-content Craters Game Pathology
Metrics: BD-rate PSNR MS-SSIM bpp transmitted parameters MFLOPs/pixel decoding time training speed
Baselines: Full-model finetune DANICE VTM-22.0 intra coding Cheng2020 TCM MLIC++
Builds on: LoRA DANICE ELIC
Closest papers by meaning
- Balanced Rate-Distortion Optimization in Learned Image Compression (CVPR 2025 · similarity 0.88)
- Frequency-Aware Transformer for Learned Image Compression (ICLR 2024 · similarity 0.87)
- On Disentangled Training for Nonlinear Transform in Learned Image Compression (ICLR 2025 · similarity 0.87)
- Test-time Adaptation for Image Compression with Distribution Regularization (ICLR 2025 · similarity 0.86)
- Causal Context Adjustment Loss for Learned Image Compression (NeurIPS 2024 · similarity 0.86)
- GLIC: General Format Learned Image Compression (AAAI 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.