ResMatching: Noise-Resilient Computational Super-Resolution via Guided Conditional Flow Matching

Anirban Ray1,2, Vera Galinova1, Florian Jug1
1Human Technopole, Milan, Italy
2Technische Universität Dresden, Germany
arXiv GitHub

Abstract

Computational Super-Resolution (CSR) in fluorescence microscopy aims to recover unobserved spatial frequencies by leveraging strong image priors. We introduce ResMatching, a guided conditional flow-matching framework that learns improved data-priors for CSR. Evaluated on four BioSR structures against seven baselines, ResMatching consistently achieves the best trade-off between fidelity and perceptual realism, especially under high noise. Moreover, it enables calibrated posterior sampling, providing pixel-wise uncertainty maps to identify unreliable predictions.

Key contributions: (1) A flow-matching framework for noise-resilient super-resolution, (2) Superior performance on diverse microscopy imaging tasks, (3) Efficient sampling from learned distributions for uncertainty calibration.

Results

Distortion and Perception Performance

Interactive Comparison Tool

Explore our results interactively. Select different datasets and image IDs, adjust crop regions, and compare ResMatching with baseline methods using the curtain tool below.

Dataset:
Image ID:
Mode:
Crop Size: 128
Position:
X: Y:

Image Comparison

Input Frame

Click and drag

Input Crop

Cropped Input Region

Ground Truth

Full GT Image

GT Crop

Cropped GT Region

Predictions

Curtain Comparison

Compare Method:
Curtain Position: 50%
Zoom: 100%
Input
Ground Truth

Citation

If you find this work useful in your research, please consider citing: 

@article{resmatching2025,
  title={ResMatching: Noise-Resilient Computational Super-Resolution via Guided Conditional Flow Matching},
  author={Anirban Ray and Vera Galinova and Florian Jug},
  journal={arXiv preprint arXiv:2510.26601},
  year={2025}
}