Research map / Image generation / Image restoration
Test-Time Degradation Adaptation for Open-Set Image Restoration
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 TAO, a test-time degradation adaptation framework that guides a fixed pretrained diffusion model with a per-sample adversarially trained adapter and stage-wise guidance to restore images with unseen degradations.
Problem
Open-set image restoration: restoring images with unknown degradations not seen during pretraining, which is framed as unidentified distribution shift between test and training data.
Approach
Uses a degradation-agnostic pretrained unconditional diffusion model as a frozen generic prior and, at test time, optimizes a small four-layer convolutional adapter per denoising step to align generated images to the test degraded image; adapter-based gradients then guide the denoising, with guidance losses changed across three denoising stages.
Key results
- On HSTS dehazing, TAO reaches 22.29 PSNR / 0.877 SSIM, the best among zero-shot methods and above AOD and CAP, but 1.67 PSNR below DEN.
- On LOL low-light enhancement, TAO gets 17.10 PSNR / 0.804 SSIM, best SSIM among zero-shot methods; only MBLLEN beats it on both metrics.
- On Kodak24 denoising (sigma=30), TAO gets 28.49 PSNR / 0.830 SSIM, comparable to BM3D (28.63 / 0.795) and N2S (28.71 / 0.864).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TAO instead of DEN | Image dehazing on HSTS · HSTS | PSNR: 23.96 -> 22.29 | Hurt |
| Use TAO instead of DEN | Image dehazing on HSTS · HSTS | SSIM: 0.902 -> 0.877 | Hurt |
| Use TAO instead of MBLLEN | Low-light enhancement · LOL test | PSNR: 18.03 -> 17.10 | Hurt |
| Use TAO instead of MBLLEN | Low-light enhancement · LOL test | SSIM: 0.809 -> 0.804 | Hurt |
| Use TAO instead of N2S | Image denoising · Kodak24 | PSNR: 28.71 -> 28.49 | Hurt |
| Use TAO instead of N2S | Image denoising · Kodak24 | SSIM: 0.864 -> 0.830 | Hurt |
| Replace VGAN_S with WGAN for domain alignment (ablation) | TAO image dehazing · HSTS | PSNR: 22.29 -> 13.61 | Hurt |
| Replace VGAN_S with WGAN for domain alignment (ablation) | TAO image dehazing · HSTS | SSIM: 0.877 -> 0.619 | Hurt |
| Use VGAN_B, LSGAN, WGAN_Softplus, or Hinge instead of VGAN_S (ablation) | TAO image dehazing · HSTS | PSNR: 22.29; 22.06; 20.76; 21.47; 20.44 | Mixed |
| Use the same MSE guidance throughout denoising instead of AIR's staged guidance (ablation) | TAO image dehazing · HSTS | PSNR: 22.29 -> 20.49 | Hurt |
| Add the first guidance stage (ablation) | TAO image dehazing · HSTS | PSNR: 21.43 -> 22.29 | Helped |
| Add the second guidance stage (ablation) | TAO image dehazing · HSTS | PSNR: 21.06 -> 22.29 | Helped |
Surprising findings
- Adversarial domain-alignment variant WGAN fails badly (13.61 PSNR) while other GAN losses work, showing sensitivity to the adversarial loss choice.
- Removing the first (global-content MSE) stage barely hurts dehazing SSIM and yields 21.43 PSNR, while using a uniform MSE guidance drops PSNR to 20.49.
Limitations
- Evaluated only on three degradation types with small test subsets (HSTS 20 images, LOL 15 pairs, Kodak 24 images), and results vary with per-task loss weights and guidance scale tuned per degradation.
- Stage boundaries (999-700, 700-50, 50-0) are chosen heuristically and tuned on the test set, and the method does not outperform the best supervised methods on all metrics.
- Per-sample optimization at every denoising step is computationally costly, and no runtime is reported.
Experimental setup
Datasets: ImageNet HSTS (RESIDE) LOL Kodak24
Benchmarks: HSTS LOL Kodak24
Metrics: PSNR SSIM
Baselines: DehazeNet MSCNN AOD-Net CAP DCP BCCR GRM NLD DDIP YOLY LightenNet Retinex-Net
Builds on: denoising diffusion probabilistic models classifier guidance Zero-DCE Deep Image Prior
Closest papers by meaning
- GenDeg: Diffusion-based Degradation Synthesis for Generalizable All-In-One Image Restoration (CVPR 2025 · similarity 0.88)
- Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · similarity 0.88)
- Visual-Instructed Degradation Diffusion for All-in-One Image Restoration (CVPR 2025 · similarity 0.87)
- Learning Where and What to Restore for Composite Image Restoration (NeurIPS 2026 · similarity 0.87)
- Parameter Efficient Adaptation for Image Restoration with Heterogeneous Mixture-of-Experts (NeurIPS 2024 · similarity 0.87)
- UP-Restorer: When Unrolling Meets Prompts for Unified Image Restoration (AAAI 2025 · similarity 0.87)
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.