Research map / Image generation / Image restoration
IR-CM: The Fast and General-purpose Image Restoration Method Based on Consistency 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 IR-CM, a consistency-training method on the IR-SDE mean-reverting SDE that enables one- or two-step general image restoration, using an origin-estimated consistency function, origin-guided loss, and a linear-nonlinear decoupling two-stage training.
Problem
Diffusion-based image restoration (e.g., IR-SDE) requires many sampling steps, which is too slow for real-time use, and consistency models usually depend on pre-trained diffusion checkpoints or underperform when trained from scratch.
Approach
Applies consistency training (and distillation) to the IR-SDE mean-reverting SDE with an origin-estimated consistency function that predicts x0 directly, an origin-guided loss to avoid trivial solutions, and a two-stage linear-fitting (x0 = mu, pure noising) then nonlinear-fitting training that decouples the degradation's linear and nonlinear parts.
Key results
- On LOLv2, IR-CM-CT with two steps reaches 30.20 PSNR / 0.95 SSIM / 0.021 LPIPS, outperforming DiffLL and GlobalDiff and the IR-SDE baseline (27.05 PSNR).
- On R100H deraining, one-step IR-CM-CT attains 30.47 PSNR / 0.92 SSIM / 0.016 LPIPS, versus IR-SDE's 31.65 / 0.90 / 0.047 at 100 NFE; the authors claim best SSIM and LPIPS.
- Ablation on LOLv2 shows the full combination (OECF + origin-guided loss + linear-nonlinear decoupling) improves LPIPS from 0.144 (baseline) to 0.027 at one step.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use two-step rather than one-step sampling | IR-CM-CT · R100H | LPIPS: 0.016 -> 0.015 | Helped |
| Use two-step rather than one-step sampling | IR-CM-CT · R100L | LPIPS: 0.009 -> 0.005 | Helped |
| Use two-step rather than one-step sampling | IR-CM · Raindrop | LPIPS: 0.043 -> 0.041 | Helped |
| Use two-step rather than one-step sampling | IR-CM-CT · GoPro | LPIPS: 0.013 -> 0.012 | Helped |
| Use two-step rather than one-step sampling | IR-CM-CT · McMaster, σ = 25 | LPIPS: 0.048 -> 0.027 | Helped |
| Use two-step rather than one-step sampling | IR-CM-CT · LOLv2 | LPIPS: 0.027 -> 0.021 | Helped |
| Add OECF (ablation) | IR-CM-CT · LOLv2 | LPIPS: 0.072 -> 0.027 | Helped |
| Add OGL (ablation) | IR-CM-CT · LOLv2 | LPIPS: 0.101 -> 0.027 | Helped |
| Add LLDT (ablation) | IR-CM-CT · LOLv2 | LPIPS: 0.059 -> 0.027 | Helped |
| Set origin-guided loss weight λOG to 0.8 rather than 0 (ablation) | IR-CM-CT · LOLv2 | LPIPS: 0.136 -> 0.027 | Helped |
| Increase λOG from 0.8 to 1.0 (ablation) | IR-CM-CT · LOLv2 | LPIPS: 0.027 -> 0.042 | Hurt |
| Use linear-fitting stage only rather than two-stage training (ablation) | IR-CM-CT · LOLv2 | LPIPS: 0.02693 -> 0.02803 | Hurt |
Surprising findings
- Consistency distillation of IR-SDE (IR-CM-CD) underperforms consistency training initialized from the IR-SDE checkpoint on several tasks, e.g., LOLv2 PSNR 26.33 vs 27.61 for CT.
- Linear-fitting stage alone yields one-step results close to the full two-stage model on LOLv2 (27.24 vs 27.61 PSNR).
Limitations
- Not suitable for unconditional generation; applicable only to conditional restoration tasks.
- Performance on PSNR is sometimes below baselines (e.g., deraining and deblurring PSNR lower than MAXIM/Restormer), with the method prioritizing perceptual metrics.
- Origin-guided loss weight is sensitive; too low causes instability and too high degrades to a CNN-baseline behavior.
Experimental setup
Datasets: LOLv2 GoPro R100L R100H Raindrop McMaster DIV2K Flickr2K BSD500
Benchmarks: LOLv2 GoPro R100L R100H Raindrop McMaster
Metrics: PSNR SSIM LPIPS NFE
Baselines: IR-SDE MAXIM Restormer JORDER PReNet DiffLL GlobalDiff Retinexformer LLFlow DiffIR
Builds on: IR-SDE consistency models consistency distillation consistency training
Closest papers by meaning
- IRPO: Boosting Image Restoration via Post-training GRPO (NeurIPS 2026 · similarity 0.88)
- IRBridge: Solving Image Restoration Bridge with Pre-trained Generative Diffusion Models (ICML 2025 · similarity 0.87)
- Reconciling Stochastic and Deterministic Strategies for Zero-shot Image Restoration using Diffusion Model in Dual (CVPR 2025 · similarity 0.87)
- Blind Image Restoration via Fast Diffusion Inversion (NeurIPS 2024 · similarity 0.87)
- CoDe: An Explicit Content Decoupling Framework for Image Restoration (CVPR 2024 · similarity 0.87)
- Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · 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.