Research map / Image generation / Image restoration
Reversing Flow for 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 ResFlow, an image restoration framework that models degradation as a deterministic continuous normalizing flow with an auxiliary variable that disambiguates the HQ uncertainty, trained by velocity matching and sampled in fewer than four steps.
Problem
Image restoration is ill-posed because degradation erases information, so one LQ image maps to many HQ images; existing diffusion/SDE-based restoration treats degradation as stochastic, which is inefficient and complex.
Approach
Model degradation as a deterministic flow between HQ and LQ images, augmented with an auxiliary variable y_t that couples with the uncertainty scope and disambiguates velocity, using an entropy-preserving schedule and velocity matching with a time-weighted loss.
Key results
- On SIDD real denoising, ResFlow reaches 42.26 dB PSNR, 2.24 dB above Restormer.
- On Outdoor-Rain deraining, ResFlow improves PSNR by 1.83 dB over NAFNet (32.82 vs 30.99 dB for the next best listed DTPM-4).
- On DPDD combined defocus deblurring, ResFlow achieves 26.96 dB PSNR, 0.78 dB above FocalNet, using four sampling steps.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ResFlow instead of DTPM-4 | image restoration · Snow100K | PSNR: +0.86dB | Helped |
| Use ResFlow instead of DTPM-4 | image restoration · Outdoor-Rain | PSNR: +1.83dB | Helped |
| Use ResFlow instead of FocalNet | image restoration · NH-HAZE | PSNR: +1.01dB | Helped |
| Use ResFlow instead of Restormer | image restoration · SIDD | PSNR: +2.24dB | Helped |
| Use ResFlow instead of SCD-Former | image restoration · LHP | PSNR: +0.20dB | Helped |
| Use ResFlow instead of NAFNet | image restoration · RealSnow | PSNR: +0.21dB | Helped |
| Use ResFlow instead of FocalNet | image restoration · DPDD combined | PSNR: +0.78dB | Helped |
| Use entropy-preserving rather than constant degradation schedule (ablation) | ResFlow · Four real-world and synthetic datasets (average) | PSNR: 23.81 -> 25.51 | Helped |
| Use Gaussian rather than constant auxiliary variable (ablation) | ResFlow · Four real-world and synthetic datasets (average) | PSNR: 23.77 -> 25.51 | Helped |
| Use adapter rather than addition for auxiliary-variable injection (ablation) | ResFlow · Four real-world and synthetic datasets (average) | PSNR: 24.04 -> 25.51 | Helped |
| Use adapter rather than channel-wise concatenation for auxiliary-variable injection (ablation) | ResFlow · Four real-world and synthetic datasets (average) | PSNR: 24.94 -> 25.51 | Helped |
| Use ResFlow instead of DAGN | JPEG artifact removal · LIVE1 (JPEG QF=10) | PSNR: 27.95 -> 27.95 | No clear effect |
Limitations
- Restoration results for the auxiliary variable rely on a learned coupling and sampling schedule that the paper does not fully characterize
- Evaluated only on image restoration tasks; extension to more complex degradations and video is left to future work
Experimental setup
Datasets: Snow100K RealSnow Outdoor-Rain LHP Dense-Haze NH-HAZE SIDD DPDD LIVE1 BSD500
Benchmarks: Snow100K RealSnow Outdoor-Rain LHP Dense-Haze NH-HAZE SIDD DPDD LIVE1 BSD500
Metrics: PSNR SSIM LPIPS MAE
Baselines: Restormer NAFNet FocalNet DTPM SCD-Former MPRNet FBCNN DAGN SwinIR IPT
Builds on: continuous normalizing flow flow matching DDPM InDI
Closest papers by meaning
- Posterior-Mean Rectified Flow: Towards Minimum MSE Photo-Realistic Image Restoration (ICLR 2025 · similarity 0.89)
- Resfusion: Denoising Diffusion Probabilistic Models for Image Restoration Based on Prior Residual Noise (NeurIPS 2024 · similarity 0.88)
- PnP-Flow: Plug-and-Play Image Restoration with Flow Matching (ICLR 2025 · similarity 0.87)
- Visual-Instructed Degradation Diffusion for All-in-One Image Restoration (CVPR 2025 · similarity 0.87)
- Continuous Expert Assembly: Instance-Conditioned Low-Rank Residuals for All-in-One Image Restoration (NeurIPS 2026 · similarity 0.87)
- Residual-Conditioned Optimal Transport: Towards Structure-Preserving Unpaired and Paired Image Restoration (ICML 2024 · similarity 0.87)
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.