Research map / Image generation / Image restoration
GenDeg: Diffusion-based Degradation Synthesis for Generalizable All-In-One Image Restoration
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 GenDeg, a degradation- and intensity-conditioned latent diffusion model that synthesizes six types of degradations on clean images, and uses 550k generated samples to build the GenDS dataset (750k+), which improves out-of-distribution generalization of all-in-one restoration models.
Problem
All-in-one image restoration models generalize poorly to out-of-distribution degradations because existing paired datasets are small, scene-limited, and have narrow degradation patterns, while collecting real paired data is infeasible.
Approach
Fine-tune an InstructPix2Pix-style Stable Diffusion model conditioned on clean image, text prompt, and the mean and standard deviation of the degradation map (binned and embedded), plus a structure correction module to fix VAE-induced misalignment, to generate many diverse paired degradations on clean scenes.
Key results
- Training PromptIR on GenDS improves OoD LPIPS on REVIDE from 0.262 to 0.212 and on O-Haze from 0.333 to 0.160 (LPIPS/FID table).
- Mean OoD LPIPS for models trained on existing data is 0.211 vs 0.1694 with GenDS in the Swin model analysis (Table 4).
- Within-distribution performance improves notably for haze, low-light and raindrop, e.g. PromptIR haze LPIPS 0.309 to 0.210 with GenDS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train NAFNet with GenDS instead of existing restoration datasets alone | NAFNet · REVIDE (Haze, real) | LPIPS: 0.211 -> 0.151 | Helped |
| Train NAFNet with GenDS instead of existing restoration datasets alone | NAFNet · O-Haze (Haze, real) | LPIPS: 0.183 -> 0.143 | Helped |
| Train NAFNet with GenDS instead of existing restoration datasets alone | NAFNet · RainDS (Rain, synthetic) | LPIPS: 0.107 -> 0.100 | Helped |
| Train NAFNet with GenDS instead of existing restoration datasets alone | NAFNet · LHP (Rain, real) | LPIPS: 0.200 -> 0.180 | Helped |
| Train NAFNet with GenDS instead of existing restoration datasets alone | NAFNet · RSVD (Snow, synthetic) | LPIPS: 0.131 -> 0.110 | Helped |
| Train NAFNet with GenDS instead of existing restoration datasets alone | NAFNet · GoPro (Motion Blur, real) | LPIPS: 0.155 -> 0.149 | Helped |
| Train NAFNet with GenDS instead of existing restoration datasets alone | NAFNet · LOLv1 (Low-light, real) | LPIPS: 0.167 -> 0.147 | Helped |
| Train NAFNet with GenDS instead of existing restoration datasets alone | NAFNet · SICE (Low-light, real) | LPIPS: 0.304 -> 0.278 | Helped |
| Train NAFNet with GenDS instead of existing restoration datasets alone | NAFNet · RainDS (Raindrop, real) | LPIPS: 0.178 -> 0.170 | Helped |
| Train NAFNet with GenDS instead of existing restoration datasets alone | NAFNet · GoPro (Motion Blur, real) | FID: 28.2 -> 28.7 | Hurt |
Surprising findings
- Training solely on GenDeg synthetic data lowers within-distribution performance due to a domain gap from alignment discrepancies, though mixing it with existing data helps.
- Motion blur OoD performance stays nearly unchanged after adding synthetic data, since that degradation already has abundant real data.
Limitations
- Scaling beyond 500k synthetic samples was not explored due to limited computational resources.
- Structure correction module blurs rain streaks, snow and low-light details, so it is omitted for those degradations.
- Synthetic-only training lowers in-distribution performance due to alignment/domain gap.
Experimental setup
Datasets: RESIDE Snow100K Rain13K GoPro REVIDE O-Haze RainDS LHP RSVD LOLv1 SICE GenDS
Benchmarks: REVIDE O-Haze RainDS LHP RSVD GoPro LOLv1 SICE
Metrics: LPIPS FID PSNR SSIM
Baselines: NAFNet PromptIR DA-CLIP Diff-Plugin DiffUIR InstructIR AutoDIR
Builds on: InstructPix2Pix Stable Diffusion AutoDIR structure correction module DA-CLIP Diff-Plugin
Closest papers by meaning
- DGSolver: Diffusion Generalist Solver with Universal Posterior Sampling for Image Restoration (NeurIPS 2025 · similarity 0.90)
- DreamClear: High-Capacity Real-World Image Restoration with Privacy-Safe Dataset Curation (NeurIPS 2024 · similarity 0.89)
- Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · similarity 0.88)
- UP-Restorer: When Unrolling Meets Prompts for Unified Image Restoration (AAAI 2025 · similarity 0.88)
- Restoration by Generation with Constrained Priors (CVPR 2024 · similarity 0.88)
- Unsupervised Diffusion-Based Degradation Modeling for Real-World Super-Resolution (AAAI 2025 · similarity 0.88)
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.