Research map / Image generation / Image restoration
DDR: Exploiting Deep Degradation Response as Flexible Image Descriptor
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 Deep Degradation Response (DDR), measuring how CLIP image features shift under text-driven feature-domain degradation, and uses it as a zero-shot opinion-unaware IQA metric and as an unsupervised loss for deblurring and SISR.
Problem
Image descriptors from deep features usually need human-labeled quality scores and target fixed attributes; the paper asks whether the response of deep features to degradation can serve as a flexible, training-free image descriptor for quality assessment and restoration.
Approach
Degradation is simulated in CLIP feature space by adding an AdaIN-stylized text-derived degradation direction (from paired degraded/clean prompts) to the image feature, and the cosine distance between original and degraded features defines DDR; averaging DDR over several degradation types gives a blind quality score and an extra loss term for restoration training.
Key results
- On OU-BIQA, DDR achieves SRCC 0.8289 on CSIQ and 0.6613 on LIVE in-the-wild, the best among the compared opinion-unaware methods on most datasets.
- On GoPro with NAFNet, PSNR+DDR reaches 33.3427 dB vs 33.1717 dB for PSNR-only loss; on RealBlur with Restormer, 32.1759 dB vs 31.9816 dB.
- On real-world SISR with NAFNet, PSNR+DDR raises PSNR from 27.0856 to 27.3121 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DDR instead of the strongest listed baseline | Zero-shot BIQA · CSIQ | SRCC: 0.8045 -> 0.8289 | Helped |
| Use DDR instead of the strongest listed baseline | Zero-shot BIQA · LIVE | SRCC: 0.9378 -> 0.8793 | Hurt |
| Use DDR instead of the strongest listed baseline | Zero-shot BIQA · TID2013 | SRCC: 0.4938 -> 0.5844 | Helped |
| Use DDR instead of the strongest listed baseline | Zero-shot BIQA · KADID | SRCC: 0.5406 -> 0.5968 | Helped |
| Use DDR instead of the strongest listed baseline | Zero-shot BIQA · KonIQ | SRCC: 0.6401 -> 0.6455 | Helped |
| Use DDR instead of the strongest listed baseline | Zero-shot BIQA · LIVEitw | SRCC: 0.5060 -> 0.6613 | Helped |
| Use DDR instead of the strongest listed baseline | Zero-shot BIQA · CID2013 | SRCC: 0.7698 -> 0.8009 | Helped |
| Use DDR instead of the strongest listed baseline | Zero-shot BIQA · SPAQ | SRCC: 0.7084 -> 0.7249 | Helped |
| Add DDR to the PSNR training objective | NAFNet image deblurring · GoPro | PSNR: 33.1717 -> 33.3427 | Helped |
| Add DDR to the PSNR training objective | NAFNet image deblurring · RealBlur | PSNR: 30.6373 -> 30.7982 | Helped |
| Add DDR to the PSNR training objective | Restormer image deblurring · GoPro | PSNR: 33.3398 -> 33.4946 | Helped |
| Add DDR to the PSNR training objective | Restormer image deblurring · RealBlur | PSNR: 31.9816 -> 32.1759 | Helped |
Surprising findings
- Larger CLIP backbones (RN50x16, RN50x64) did not improve deblurring performance, suggesting larger visual models do not necessarily better capture low-level texture.
- Removing the AdaIN-style adaptation drops CSIQ SRCC from 0.8289 to 0.6074, showing feature-domain fusion without adaptation is much weaker.
Limitations
- CLIP visual features focus on high-level content and may poorly capture low-level degradation.
- Suitable degradation prompts vary across downstream tasks and are chosen by hand.
- Restoration experiments are run at 128x128 training resolution with a single GPU and limited settings.
Experimental setup
Datasets: LIVEitw CSIQ TID2013 KADID KonIQ LIVE CID2013 SPAQ GoPro RealBlur RealSR City100
Benchmarks: CSIQ LIVE TID2013 KADID KonIQ LIVEitw CID2013 SPAQ GoPro RealBlur RealSR+City100
Metrics: SRCC PSNR SSIM
Baselines: NIQE QAC PIQE LPSI ILNIQE dipIQ SNP-NIQE NPQI ContentSep LPIPS CTX PDL
Builds on: CLIP AdaIN manifold data augmentation (Moon et al.) NAFNet Restormer
Closest papers by meaning
- Unsupervised Degradation Representation Aware Transform for Real-World Blind Image Super-Resolution (AAAI 2025 · similarity 0.89)
- Degradation-Aware Feature Perturbation for All-in-One Image Restoration (CVPR 2025 · similarity 0.88)
- Vision-Language Gradient Descent-driven All-in-One Deep Unfolding Networks (CVPR 2025 · similarity 0.87)
- Controlling Vision-Language Models for Multi-Task Image Restoration (ICLR 2024 · similarity 0.87)
- Universal Image Restoration Pre-training via Degradation Classification (ICLR 2025 · similarity 0.86)
- Unsupervised Diffusion-Based Degradation Modeling for Real-World Super-Resolution (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.