Research map / Image generation / Image restoration

Reti-Diff: Illumination Degradation Image Restoration with Retinex-based Latent Diffusion Model

ICLR 2025 · Image generation › Image restoration · approach: Diffusion-based universal restoration · 10 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 Reti-Diff, a latent diffusion model that generates Retinex reflectance and illumination priors to guide a transformer for illumination degradation image restoration across low-light, underwater, and backlit enhancement tasks.

Problem

Illumination degradation image restoration (IDIR) with diffusion models is computationally expensive when run at image level and suffers from pixel misalignment in restored details and local consistency.

Approach

Runs the diffusion process in a compact latent space to predict Retinex reflectance and illumination priors (via two Siamese RLDMs) from the low-quality input, and uses these priors to guide a Retinex-guided transformer (RGformer) that decomposes features into reflectance and illumination parts for refinement.

Key results

• On LOL-v1, Reti-Diff reaches 25.35 PSNR / 0.866 SSIM, and 13.2% better than the second-best method on average for low-light enhancement per the paper.
• On the ExDark low-light detection task, Reti-Diff gets 75.8 mAP versus 72.5 for Retformer, a 4.72% gain.
• Reti-Diff runs at 12.27 FPS with 156.55 G MACs and 26.11M parameters, versus 4.25 FPS for Diff-Retinex.

Measured effects reported in the paper

Surprising findings

• Using RLDM with only T=4 diffusion steps is reported as sufficient for stable priors, with performance converging quickly.

Limitations

• Produces blurred texture details under severe illumination degradation (failure cases).
• Texture details are lost during illumination recovery; the paper proposes frequency-domain priors as a remedy.
• Only IDIR tasks with low-light, underwater, and backlit degradations were studied; other degradations like haze and motion blur are future work.

Experimental setup

Datasets: LOL-v1 LOL-v2-real LOL-v2-synthetic SID UIEB LSUI BAID DICM LIME MEF NPE VV

Benchmarks: LOL-v1 LOL-v2-real LOL-v2-synthetic SID UIEB LSUI BAID DICM LIME MEF NPE VV

Metrics: PSNR SSIM FID BIQE UCIQE UIQM LPIPS PI NIQE mAP

Baselines: Diff-Retinex PyDiff GSAD Retformer SNR-Net CLIP-LIT DiffIR Restormer MIRNet URetinex DCC-Net PUGAN

Builds on: Retformer Diff-Retinex DiffIR CIDNet

Closest papers by meaning

• Diff-Shadow: Global-guided Diffusion Model for Shadow Removal (AAAI 2025 · similarity 0.87)
• DGSolver: Diffusion Generalist Solver with Universal Posterior Sampling for Image Restoration (NeurIPS 2025 · similarity 0.86)
• IniRetinex: Rethinking Retinex-type Low-Light Image Enhancer via Initialization Perspective (AAAI 2025 · similarity 0.86)
• Selective Hourglass Mapping for Universal Image Restoration Based on Diffusion Model (CVPR 2024 · similarity 0.85)
• DreamClear: High-Capacity Real-World Image Restoration with Privacy-Safe Dataset Curation (NeurIPS 2024 · similarity 0.85)
• A Modular Conditional Diffusion Framework for Image Reconstruction (NeurIPS 2024 · similarity 0.85)