Research map / Label-efficient and robust learning / Domain adaptation and generalization

Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration

ICLR 2025 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Geometric alignment and synthetic-to-real adaptation · 0 citations (OpenAlex)

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 "denoising as adaptation", a noise-space domain adaptation method where a jointly trained diffusion denoiser, conditioned on restored synthetic and real outputs, supplies a diffusion loss that aligns a U-Net restoration network's real-world outputs to the clean distribution, with channel shuffling and residual-swapping contrastive learning to prevent shortcuts.

Problem

Learning-based image restoration trained on synthetic degradations generalizes poorly to real-world images due to the domain gap; existing domain adaptation methods (feature- or pixel-space) are unstable or ill-suited to low-level restoration.

Approach

Use a diffusion model's multi-step denoising loss, conditioned on restored synthetic and real outputs, as a domain adaptation signal in noise space; the restoration network is pushed to provide conditions that lower the diffusion loss, aligning both domains to a clean target distribution. Channel shuffling and residual-swapping contrastive learning prevent shortcut learning, and the diffusion mode

Key results

• On SIDD denoising, the method reaches 34.71 dB PSNR / 0.9202 SSIM versus 26.58 dB for the vanilla U-Net trained on synthetic data only.
• On SPA deraining it reaches 34.39 dB PSNR, slightly above Restormer (34.17 dB), and on RealBlur-J deblurring 26.46 dB PSNR versus 26.27 dB for vanilla.
• With a deeper U-Net (Ours*), SIDD PSNR improves to 35.52 dB, and the method also improves DnCNN, Restormer and SwinIR on SIDD.

Measured effects reported in the paper

Surprising findings

• Excluding real data from the diffusion conditions drops performance to roughly the vanilla level, indicating both domains are needed.
• Training with only a diffusion loss on real data (no synthetic conditions) still yields 32.60 dB on SIDD, though lower than the joint setting.

Limitations

• Gains are smaller for deblurring, since low-frequency artifacts are less sensitive to the diffusion loss, which the authors attribute to the high-frequency nature of diffusion denoising.
• Fails on challenging real-world degradations such as strong or out-of-distribution noise (failure cases in appendix).
• Training requires paired synthetic data with ground truth plus real degraded images, and an extra diffusion network during training.

Experimental setup

Datasets: DIV2K Flickr2K WED BSD SIDD Rain13K SPA GoPro RealBlur-J MS-COCO

Benchmarks: SIDD SPA RealBlur-J DND Real-Internet

Metrics: PSNR SSIM LPIPS

Baselines: Vanilla U-Net DANN DSN PixelDA CyCADA Ne2Ne MaskedD NLCL SelfDeblur VDIP Restormer C2N

Builds on: DDPM DANN CyCADA Uformer

Closest papers by meaning

• Diffusion Priors for Variational Likelihood Estimation and Image Denoising (NeurIPS 2024 · similarity 0.90)
• Residual Denoising Diffusion Models (CVPR 2024 · similarity 0.90)
• A Variational Perspective on Solving Inverse Problems with Diffusion Models (ICLR 2024 · similarity 0.90)
• Restoration by Generation with Constrained Priors (CVPR 2024 · similarity 0.89)
• RestoreGrad: Signal Restoration Using Conditional Denoising Diffusion Models with Jointly Learned Prior (ICML 2025 · similarity 0.89)
• Resfusion: Denoising Diffusion Probabilistic Models for Image Restoration Based on Prior Residual Noise (NeurIPS 2024 · similarity 0.89)