Research map / Image generation / Image restoration
Selective Hourglass Mapping for Universal Image Restoration Based on Diffusion Model
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 DiffUIR, a conditional diffusion model for universal image restoration that maps all degradations to a shared impure Gaussian distribution via a shared distribution term, then recovers task-specific outputs with strong condition guidance.
Problem
Universal image restoration must handle different degradation distributions in one model; existing multi-partite mapping methods (task-specific prompts or encoders) learn each distribution separately and ignore information shared across tasks.
Approach
Combines RDDM-style explicit and implicit condition injection with a shared distribution term (SDT) added to the diffusion forward and reverse processes, gradually mapping different degraded distributions to one shared impure Gaussian (selective hourglass mapping) and recovering task-specific clean images.
Key results
- DiffUIR-L (36.26M params) reaches 31.03 dB PSNR on deraining, 25.12 dB on low-light, 32.65 dB on desnowing (Snow100K-S), 32.94 dB on dehazing and 29.17 dB on deblurring in the universal setting.
- DiffUIR-T with only 0.89M parameters and 1.72 GFLOPs achieves 28.60 dB deraining and 27.25 dB deblurring, outperforming several larger universal methods such as AirNet and Prompt-IR on several tasks.
- On unknown-task under-display camera restoration (TOLED), DiffUIR reaches 29.55 dB PSNR versus 27.91 dB for IDR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DiffUIR-L instead of Prompt-IR | universal image restoration · Deraining (5sets) | PSNR: 29.56 -> 31.03 | Helped |
| use DiffUIR-L instead of DA-CLIP | universal image restoration · Enhancement | PSNR: 24.17 -> 25.12 | Helped |
| use DiffUIR-L instead of Prompt-IR | universal image restoration · Desnowing (2sets) | PSNR: 31.98 -> 32.65 | Helped |
| use DiffUIR-L instead of Prompt-IR | universal image restoration · Dehazing | PSNR: 32.02 -> 32.94 | Helped |
| use DiffUIR-L instead of Prompt-IR | universal image restoration · Deblurring | PSNR: 27.21 -> 29.17 | Helped |
| add SDT to DiffUIR (ablation) | DiffUIR · five restoration tasks | PSNR: 30.34 -> 31.03; 22.99 -> 25.12; 32.34 -> 32.65; 32.47 -> 32.94; 28.50 -> 29.17 | Helped |
| retain explicit condition and SDT (ablation) | DiffUIR · five restoration tasks | PSNR: 30.63 -> 31.03; 24.28 -> 25.12; 32.40 -> 32.65; 32.35 -> 32.94; 28.62 -> 29.17 | Helped |
| retain implicit condition and SDT (ablation) | DiffUIR · five restoration tasks | PSNR: 22.93 -> 31.03; 20.93 -> 25.12; 23.96 -> 32.65; 24.26 -> 32.94; 27.08 -> 29.17 | Helped |
| set shared distribution coefficient to 0.1 rather than 0.9 (ablation) | DiffUIR · Deraining (5sets) | PSNR: 31.03 -> 30.70 | Hurt |
| set shared distribution coefficient to 0.25 rather than 0.9 (ablation) | DiffUIR · Deraining (5sets) | PSNR: 31.03 -> 30.89 | Hurt |
| set shared distribution coefficient to 0.5 rather than 0.9 (ablation) | DiffUIR · Deraining (5sets) | PSNR: 31.03 -> 31.01 | Hurt |
| set shared distribution coefficient to 0.75 rather than 0.9 (ablation) | DiffUIR · Deraining (5sets) | PSNR: 31.03 -> 30.77 | Hurt |
Surprising findings
- Mapping degradations to an impure Gaussian (keeping 10% of input) works better than a pure Gaussian endpoint; performance drops when the shared coefficient reaches 1.
- A tiny 0.89M-parameter model remains competitive with universal methods having tens to hundreds of millions of parameters.
Limitations
- Performance depends on batch size because shared distribution learning requires all tasks in the same batch.
- Diffusion sampling requires multiple steps (3 timesteps used), and performance improves with more steps.
Experimental setup
Datasets: Rain datasets (merged, from refs 17 and 60) LOL Snow100K RESIDE GoPro TOLED POLED Practical MEF NPE DICM HIDE
Benchmarks: Snow100K-S Snow100K-L LOL RESIDE outdoor GoPro merged deraining Practical MEF NPE DICM HIDE RealBlur-J
Metrics: PSNR SSIM LPIPS NIQE LOE IL-NIQE
Baselines: Restomer AirNet Painter IDR ProRes Prompt-IR DA-CLIP RDDM SwinIR MIRNet-v2 DehazeFormer MAXIM
Builds on: RDDM DDPM DDIM Restomer
Closest papers by meaning
- UP-Restorer: When Unrolling Meets Prompts for Unified Image Restoration (AAAI 2025 · similarity 0.87)
- A Modular Conditional Diffusion Framework for Image Reconstruction (NeurIPS 2024 · similarity 0.86)
- GenDeg: Diffusion-based Degradation Synthesis for Generalizable All-In-One Image Restoration (CVPR 2025 · similarity 0.86)
- Visual-Instructed Degradation Diffusion for All-in-One Image Restoration (CVPR 2025 · similarity 0.86)
- Diff-Shadow: Global-guided Diffusion Model for Shadow Removal (AAAI 2025 · similarity 0.86)
- Learning Diffusion Texture Priors for Image Restoration (CVPR 2024 · 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.