Research map / Image generation / Image restoration
AdaIR: Adaptive All-in-One Image Restoration via Frequency Mining and Modulation
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 AdaIR, a U-shaped Transformer for all-in-one image restoration that mines low- and high-frequency features guided by adaptively learned FFT masks from the degraded input and modulates them bidirectionally.
Problem
All-in-one image restoration methods operate only in the spatial domain and ignore that different degradation types (noise, rain, haze, low-light, blur) corrupt different frequency subbands, so they do not exploit frequency-specific treatment.
Approach
An Adaptive Frequency Learning Block (AFLB) inserted between decoder levels: a Mask Generation Block predicts a learnable frequency boundary from the degraded image to split features into low/high frequency via FFT, then H-L (spatial attention) and L-H (channel attention) units exchange information bidirectionally before cross-attention merges the result back into the features.
Key results
- In the three-task all-in-one setting, AdaIR averages 32.69 dB PSNR vs. PromptIR's 32.06 dB, with 2.27 dB gain on Rain100L deraining.
- In the five-task all-in-one setting, AdaIR averages 30.20 dB PSNR, 1.86 dB above IDR (28.34 dB), with over 5 dB gain on SOTS dehazing.
- Single-task dehazing on SOTS reaches 31.80 dB PSNR, 0.49 dB above PromptIR; single-task deraining on Rain100L reaches 38.90 dB, 1.86 dB above PromptIR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AdaIR instead of PromptIR | three-degradation all-in-one restoration · Rain100L | PSNR: 36.37 -> 38.64 | Helped |
| Use AdaIR instead of PromptIR | three-degradation all-in-one restoration · SOTS | PSNR: 30.58 -> 31.06 | Helped |
| Use AdaIR instead of IDR | five-degradation all-in-one restoration · SOTS dehazing | PSNR: 25.24 -> 30.53 | Helped |
| Use AdaIR instead of IDR | five-degradation all-in-one restoration · BSD68 σ=25 denoising | PSNR: 31.60 -> 31.35 | Hurt |
| Add fixed-mask frequency mining to baseline (ablation) | AdaIR · SOTS dehazing ablation | PSNR: 28.21 -> 29.79 | Helped |
| Add L-H modulation unit (ablation) | AdaIR fixed-mask variant · SOTS dehazing ablation | PSNR: 29.79 -> 30.37 | Helped |
| Add H-L modulation unit (ablation) | AdaIR fixed-mask + L-H variant · SOTS dehazing ablation | PSNR: 30.37 -> 30.52 | Helped |
| Replace fixed mask with adaptive MGB mask (ablation) | AdaIR · SOTS dehazing ablation | PSNR: 30.52 -> 31.24 | Helped |
| Replace adaptive Fourier decomposition with average pooling (ablation) | AdaIR · SOTS dehazing | PSNR: 31.24 -> 30.59 | Hurt |
| Replace adaptive Fourier decomposition with a 5×5 Gaussian filter (ablation) | AdaIR · SOTS dehazing | PSNR: 31.24 -> 30.22 | Hurt |
| Use embedding features rather than the degraded image for frequency guidance (ablation) | AdaIR · SOTS dehazing | PSNR: 30.52 -> 29.29 | Hurt |
| Use spatial attention without frequency-specific input distinction instead of FMoM (ablation) | AdaIR · SOTS dehazing | PSNR: 29.67 vs. 30.52 | Hurt |
Surprising findings
- Applying AFLBs in both encoder and decoder degrades performance (29.70 dB vs 30.52 dB for decoder-only on SOTS), contrary to the expectation that more placement helps.
- Image-level frequency decomposition from the degraded input outperforms feature-level decomposition (30.52 vs 29.29 dB on SOTS).
Limitations
- Gains on individual tasks are mixed; for example, on denoising the method is roughly on par with IDR or Restormer at some noise levels.
- Mixed-degradation and unseen-task generalization evaluated only on a few settings (desnowing, Rain100L with noise).
Experimental setup
Datasets: SOTS Rain100L BSD400 WED GoPro LOL-v1 BSD68 Urban100 Kodak24 CSD
Benchmarks: SOTS Rain100L BSD68 Urban100 GoPro LOL-v1 Kodak24 CSD
Metrics: PSNR SSIM
Baselines: PromptIR AirNet IDR MPRNet Restormer NAFNet HINet DGUNet MIRNetV2 SwinIR TAPE Transweather
Builds on: Restormer PromptIR AirNet IDR
Closest papers by meaning
- Degradation-Aware Feature Perturbation for All-in-One Image Restoration (CVPR 2025 · similarity 0.87)
- From Zero to Detail: Deconstructing Ultra-High-Definition Image Restoration from Progressive Spectral Perspective (CVPR 2025 · similarity 0.86)
- Debiased All-in-one Image Restoration with Task Uncertainty Regularization (AAAI 2025 · similarity 0.86)
- Efficient RAW Image Deblurring with Adaptive Frequency Modulation (NeurIPS 2025 · similarity 0.86)
- Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · similarity 0.86)
- GenDeg: Diffusion-based Degradation Synthesis for Generalizable All-In-One Image Restoration (CVPR 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.