Research map / Label-efficient and robust learning / Domain adaptation and generalization
IODA: Instance-Guided One-shot Domain Adaptation for Super-Resolution
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 IODA, a one-shot domain adaptation method for super-resolution that uses a single unlabeled target LR image, with image-guided CLIP direction alignment and SAM-instance-driven Alpha-CLIP masks to diversify target features.
Problem
Super-resolution networks degrade under domain gaps, but existing domain adaptation methods need many target-domain LR images, and text-guided one/zero-shot adaptation lacks the texture granularity SR needs; the target domain may consist of just one LR image.
Approach
Replace text cross-domain direction vectors with image-based CLIP direction vectors computed from LR and SR image pairs, and expand the single-image target distribution by feeding Alpha-CLIP with SAM instance masks (plus occlusion masks) to produce instance-specific features.
Key results
- On DF2K to RealSR_Canon with SAFMN, IODA raises PSNR from 24.824 to 27.461 ± 0.008 and SSIM from 0.744 to 0.790.
- Fine-tuning on all target HR samples reaches 29.875 PSNR with SAFMN, showing a gap versus the one-shot approach.
- Across backbones, IODA improves HAT on RealSR_Canon from 27.477 to 27.821 PSNR; the text-guided baseline was ineffective for SR (reported as '-').
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add image-guided domain adaptation using the LR cross-domain direction (ablation) | SAFMN · RealSR_Canon | PSNR: 24.824 -> 27.014; 2.19 | Helped |
| Add instance-guided target-domain distribution expansion (ablation) | Image-guided SAFMN adaptation · RealSR_Canon | PSNR: 27.014 -> 27.461; 0.447 | Helped |
| Replace generated source-domain SR images with real HR images (ablation) | SAFMN image-guided adaptation · RealSR_Canon | PSNR: 27.461 -> 27.157; 0.304 | Hurt |
| Use text-guided rather than image-guided domain differences (ablation) | SR domain adaptation · GTA → Cityscapes | PSNR | No clear effect |
| Add random occlusion masks (ablation) | Image-guided SAFMN adaptation · RealSR_Canon | PSNR: 27.014 -> 27.270 | Helped |
| Use random grid region-range masks instead of no diversity enhancement (ablation) | Image-guided SAFMN adaptation · RealSR_Canon | PSNR: 27.014 -> 27.259 | Helped |
| Use single random-sized region-range masks instead of no diversity enhancement (ablation) | Image-guided SAFMN adaptation · RealSR_Canon | PSNR: 27.014 -> 27.321 | Helped |
| Use instance region-range masks instead of no diversity enhancement (ablation) | Image-guided SAFMN adaptation · RealSR_Canon | PSNR: 27.014 -> 27.384 | Helped |
| Combine random occlusion masking with instance region-range masks (ablation) | Image-guided SAFMN adaptation · RealSR_Canon | PSNR: 27.014 -> 27.461 | Helped |
| Apply IODA to the SAFMN source-domain model | SAFMN · RealSR_Canon | PSNR: 24.824 -> 27.461 ± 0.008 | Helped |
| Apply IODA to the SAFMN source-domain model | SAFMN · RealSR_Nikon | PSNR: 23.980 -> 26.616 ± 0.042 | Helped |
| Apply IODA to the SAFMN source-domain model | SAFMN · GTA → Cityscapes | PSNR: 38.390 -> 38.938 ± 0.002 | Helped |
Surprising findings
- Using real HR images instead of source-network SR outputs in the direction loss lowered PSNR by 0.304, suggesting explicit source-network constraints matter.
- Text-guided domain adaptation (Style-GAN-NADA, P0DA) was reported as ineffective for SR and yielded no usable results.
Limitations
- Domain adaptation training takes over ten minutes, which is not user-friendly for single-photo use.
- Results on only a few source-target pairs and networks; ten-minute adaptation per image is costly.
- Still lags fine-tuning with all target HR samples (e.g., SAFMN 27.46 vs 29.88 PSNR).
Experimental setup
Datasets: DF2K RealSR GTA Cityscapes ACDC
Benchmarks: RealSR_Canon RealSR_Nikon GTA to Cityscapes Cityscapes to ACDC GTA to ACDC
Metrics: PSNR SSIM
Baselines: DASR DASR one-shot ZSSR Style-GAN-NADA P0DA BSRGAN Real-ESRGAN SwinIR DASR (blind SR) HAT Fine-tune
Builds on: StyleGAN-NADA Alpha-CLIP SAM MAE DASR
Closest papers by meaning
- Uncertainty-Aware Source-Free Adaptive Image Super-Resolution with Wavelet Augmentation Transformer (CVPR 2024 · similarity 0.86)
- Low-Res Leads the Way: Improving Generalization for Super-Resolution by Self-Supervised Learning (CVPR 2024 · similarity 0.86)
- Text-guided Explorable Image Super-resolution (CVPR 2024 · similarity 0.86)
- Pixel-level and Semantic-level Adjustable Super-resolution: A Dual-LoRA Approach (CVPR 2025 · similarity 0.85)
- AdaFormer: Efficient Transformer with Adaptive Token Sparsification for Image Super-resolution (AAAI 2024 · similarity 0.84)
- Learning Coupled Dictionaries from Unpaired Data for Image Super-Resolution (CVPR 2024 · similarity 0.84)
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.