Research map / Image generation / Image restoration
IRBridge: Solving Image Restoration Bridge with Pre-trained Generative Diffusion Models
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 IRBridge, a transition equation that lets a pretrained Stable Diffusion (via ControlNet) be used inside existing image restoration bridges (IR-SDE, GOUB) without training a bridge model from scratch per degradation.
Problem
Image restoration bridge models must be trained from scratch for each degradation type, which is costly and limits performance; the state distributions of pretrained diffusion models (starting from noise) and restoration bridges (starting from degraded images) mismatch.
Approach
A transition equation that converts states between two diffusion processes sharing the same endpoint distribution, using conditional independence given x0; with boundary-condition choices (critical timestep, forward/reverse transitions) the x0 dependence vanishes so no prior x0 estimate is needed, letting a conditional pretrained diffusion model (ControlNet on Stable Diffusion v1-5) act as the bri
Key results
- On image inpainting, IRBridge-GOUB reaches 30.42 PSNR and FID 3.37 versus GOUB's 28.98 PSNR and FID 4.30 on CelebA-HQ.
- On deraining, IRBridge-GOUB reaches 30.39 PSNR / 8.64 FID versus GOUB's 31.96 PSNR / 18.14 FID, so PSNR is lower for IRBridge-GOUB on that task while FID improves.
- On low-light enhancement, IRBridge-IRSDE improves over the best from-scratch bridge baseline (DiffUIR, 24.29 PSNR) with 25.52 PSNR and 0.9232 SSIM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| integrate a pretrained Stable Diffusion v1-5 prior into the image restoration bridge | IRBridge-GOUB versus GOUB · Deraining (Rain100H test) | FID: 18.14 -> 8.64 | Helped |
| integrate a pretrained Stable Diffusion v1-5 prior into the image restoration bridge | IRBridge-IRSDE versus IR-SDE · Dehazing (SOTS) | PSNR: 31.65 -> 26.85 | Helped |
| integrate a pretrained Stable Diffusion v1-5 prior into the image restoration bridge | IRBridge-GOUB versus GOUB · Desnowing (Snow100K test) | FID: 4.30 -> 10.38 | Helped |
| integrate a pretrained Stable Diffusion v1-5 prior into the image restoration bridge | IRBridge-IRSDE versus IR-SDE · Raindrop removal (RainDrop test A) | FID: 25.13 -> 48.39 | Helped |
| integrate a pretrained Stable Diffusion v1-5 prior into the image restoration bridge | IRBridge-GOUB versus GOUB · Low-light enhancement (LOL test) | PSNR: 22.98 -> 25.59 | Helped |
| integrate a pretrained Stable Diffusion v1-5 prior into the image restoration bridge | IRBridge-GOUB versus GOUB · Image inpainting (CelebA-HQ) | FID: 4.30 -> 3.37 | Helped |
| reduce inference steps from 100 to 50 (ablation) | IRBridge · Image restoration (inference-step comparison) | PSNR: 29.63 -> 29.59 | No clear effect |
| reduce inference steps from 100 to 25 (ablation) | IRBridge · Image restoration (inference-step comparison) | PSNR: 29.63 -> 29.24 | Hurt |
| remove conditional guidance (ablation) | IRBridge · Image restoration | visual detail | Hurt |
Surprising findings
- IRBridge trained only on facial data (inpainting) generalizes to indoor and outdoor scenes better than GOUB trained on the same data.
- Unconditional generation (without the degraded image as condition) retains only the layout and loses texture and detail, so conditioning is required.
Limitations
- Timestep schedules and hyperparameters are chosen empirically per task and are not theoretically optimal.
- ControlNet conditioning is a relatively coarse implementation without restoration-specific objectives.
- Iterative diffusion-style inference is slow compared with regression methods.
Experimental setup
Datasets: Rain100H RESIDE (OTS/SOTS) Snow100K RainDrop LOL CelebA-HQ RealRain-1K RTTS
Benchmarks: Rain100H RESIDE SOTS Snow100K RainDrop Testing Set A LOL CelebA-HQ inpainting RealRain-1K RTTS
Metrics: PSNR SSIM LPIPS FID MUSIQ BRISQUE NIQE
Baselines: ResShift IR-SDE GOUB RDDM DiffUIR WeatherDiff WGWS-Net GDP TAO DDRM PromptIR DCPT
Builds on: IR-SDE GOUB DDPM ControlNet Stable Diffusion
Closest papers by meaning
- Image Restoration Through Generalized Ornstein-Uhlenbeck Bridge (ICML 2024 · similarity 0.90)
- Restoration by Generation with Constrained Priors (CVPR 2024 · similarity 0.90)
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.90)
- Improving Diffusion-Based Image Restoration with Error Contraction and Error Correction (AAAI 2024 · similarity 0.89)
- A Modular Conditional Diffusion Framework for Image Reconstruction (NeurIPS 2024 · similarity 0.89)
- Reconciling Stochastic and Deterministic Strategies for Zero-shot Image Restoration using Diffusion Model in Dual (CVPR 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.