Research map / Image generation / Fast diffusion sampling
Unaligned Image Guided Denoising via Cross-modal Conditional Flow Matching
Runmin Zhang, Linshan Li, Si-Yuan Cao, Zhu Yu, Lingyu Zhu, Huaqi Zhang, Hui-liang Shen
Read the paperCheck your idea against 137,217 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 UGD-FM, a cross-modal conditional flow matching framework for denoising a target image guided by an unaligned reference image with large 2D misalignment.
Problem
Image denoising guided by a reference image when the guide is not spatially aligned with the noisy target and misalignment is large-range and two-dimensional, combined with severe target noise.
Approach
Formulates guided denoising as a progressive process built on conditional flow matching across modalities, handling large-range 2D misalignment rather than only small-baseline rectified pairs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use flow matching instead of direct regression (one step) (ablation) | UGD-FM · FANVID a=0.03 s=0.3, 1-step | PSNR: 30.27 -> 30.28 | No clear effect |
| run two solver steps instead of one in standard FM (ablation) | UGD-FM · FANVID | PSNR: 30.28 -> 30.34 | Improved |
| add stage-focused time sampling (ablation) | UGD-FM · FANVID, 2-step | PSNR: 30.34 -> 30.42 | Improved |
| add inference-consistent state construction (ablation) | UGD-FM · FANVID, 2-step | PSNR: 30.42 -> 30.58 | Improved |
| add adaptive sparse guidance propagation instead of no attention (ablation) | UGD-FM · FANVID, 2-step | PSNR: 29.10 -> 30.58 | Improved |
| use sparse propagation instead of window attention (ablation) | UGD-FM · FANVID, 2-step | PSNR: 30.03 -> 30.58 | Improved |
| use global attention for guidance matching (ablation) | UGD-FM · FANVID | PSNR: OOM | Hurt |
| add both perceptual consistency and flow consistency losses (ablation) | UGD-FM · FANVID, 2-step | PSNR: 29.78 -> 30.58 | Improved |
| add flow consistency loss only (ablation) | UGD-FM · FANVID, 2-step | PSNR: 29.78 -> 30.45 | Improved |
| run two-step rather than one-step inference | UGD-FM (Restormer) · FANVID a=0.03 s=0.3 | PSNR: 30.29 -> 30.58 | Improved |
| use guided NIR denoising instead of single-image Restormer | UGD-FM · MS2 a=0.03 s=0.3 | PSNR: 30.69 -> 32.82 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Restormer — Restormer: Efficient Transformer for High-Resolution Image Restoration (CVPR 2022) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Experimental setup
Builds on: flow matching
Closest papers by meaning
- Tuning Real-World Image Restoration at Inference: A Test-Time Scaling Paradigm for Flow Matching Models (ECCV 2026 · similarity 0.88)
- Fully Trainable and Interpretable Non-Local Sparse Models for Image Restoration (ECCV 2020 · similarity 0.87)
- Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · similarity 0.86)
- Improving Progressive Generation with Decomposable Flow Matching (NeurIPS 2025 · similarity 0.86)
- Residual Denoising Diffusion Models (CVPR 2024 · similarity 0.86)
- Diffusion Priors for Variational Likelihood Estimation and Image Denoising (NeurIPS 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.
