Research map / Image generation / Image restoration

Prompt-SID: Learning Structural Representation Prompt via Latent Diffusion for Single Image Denoising

AAAI 2025 · Image generation › Image restoration · approach: Diffusion priors for denoising and deblurring · 4 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 Prompt-SID, a self-supervised single-image denoiser that uses a latent-diffusion-generated structural representation as a prompt fused into a transformer denoiser via a structural attention module, trained on downsampled sub-image pairs with a scale replay mechanism.

Problem

Self-supervised single-image denoising via blind-spot networks or sub-image downsampling loses pixel information and destroys structural detail, and the scale gap between downsampled training images and original-resolution test images hurts performance.

Approach

A latent diffusion model (RG-Diff) generates a structural representation of the original-scale image conditioned on the sub-image representation; this representation is injected into a transformer denoiser through a structural attention module, and a scale replay branch without gradients enforces consistency at original scale.

Key results

• On SIDD validation, Prompt-SID achieves 51.55 dB / 0.992 vs. B2U 51.36 dB and NBR2NBR 51.06 dB (raw-RGB).
• On Set14 with sigma in [5,50] Gaussian noise, Prompt-SID reports 31.22 dB, 0.32 dB above the N2C baseline's 31.41? (table shows Ours 31.22 vs baseline 31.41, so the 0.32 dB claim is not consistent with the table).
• Ablation on SIDD: removing RG-Diff drops PSNR from 51.55 to 51.32 dB; removing scale replay loss drops it to 50.97 dB.

Measured effects reported in the paper

Surprising findings

• The text claims a 0.32 dB gain over the N2C baseline on Set14 for sigma in [5,50], but the reported table shows Prompt-SID at 31.22 dB versus 31.41 dB for the baseline, a discrepancy.
• Supervised N2C baseline outperforms Prompt-SID on several Gaussian and Poisson settings in the tables, despite the paper's claim of superiority in eight of twelve settings.

Limitations

• Gains over the strongest self-supervised baselines are modest (often about 0.2 dB or less on synthetic data).
• Reported results on some settings are below supervised N2C, and some text claims are inconsistent with the tables.

Experimental setup

Datasets: ILSVRC2012 SIDD-Medium SRDtrans two-photon calcium imaging dataset

Benchmarks: Kodak BSD300 Set14 SIDD benchmark SIDD validation fluorescence two-photon calcium imaging

Metrics: PSNR SSIM SNR

Baselines: N2C CBM3D N2V NBR2NBR B2U ZS-N2N DT

Builds on: Noise2Noise Noise2Void Blind2Unblind NBR2NBR ZS-N2N

Closest papers by meaning

• Prompt-tuning Latent Diffusion Models for Inverse Problems (ICML 2024 · similarity 0.87)
• Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.87)
• Diff-Prompt: Diffusion-Driven Prompt Generator with Mask Supervision (ICLR 2025 · similarity 0.86)
• Deconstructing Denoising Diffusion Models for Self-Supervised Learning (ICLR 2025 · similarity 0.86)
• FreeControl: Efficient, Training-Free Structural Control via One-Step Attention Extraction (NeurIPS 2025 · similarity 0.86)
• OSCAR: One-Step Diffusion Codec Across Multiple Bit-rates (NeurIPS 2025 · similarity 0.86)