ResMatching: Fluorescence Super-Resolution

• ResMatching is a computational super-resolution method for fluorescence microscopy that reconstructs high-resolution structures from noisy, low-resolution acquisitions.
• The approach learns a structure-aware, data-driven prior from paired LR-HR observations, balancing fidelity with perceptual realism while providing calibrated uncertainty estimates.
• By enabling posterior sampling, ResMatching produces diverse reconstructions that offer practical uncertainty quantification for enhanced analysis in noisy imaging scenarios.

ResMatching is a computational super-resolution method for fluorescence microscopy that uses guided conditional flow matching to reconstruct high-resolution (HR) structures from low-resolution (LR) acquisitions corrupted by optical blur and imaging noise. In the formulation reported for fluorescence microscopy, the method learns a structure-aware, data-driven prior directly from paired LR–HR observations, without requiring explicit knowledge of the imaging operator or noise parameters. It is presented as a conditional generative model for computational super-resolution (CSR), with posterior sampling and calibrated pixel-wise uncertainty estimates in addition to single-image reconstruction (Ray et al., 30 Oct 2025).

1. Problem formulation and ill-posedness

In the reported setting, CSR seeks to recover HR structures from LR observations of the same sample under different microscopes. The paired observations are modeled as

$x_{M_0} = H_{M_0}(s) + \eta(s),$

and

$x_{M_1} = H_{M_1}(s) + \eta(s),$

where $H_{M_0}$ and $H_{M_1}$ are unknown microscope-specific degradations and $\eta(s)$ is signal-dependent noise. Because the microscope’s optical transfer function low-pass filters signals, high-frequency content is unrecoverable from measurements alone, which makes CSR inherently ill-posed and prior-dependent (Ray et al., 30 Oct 2025).

Within this formulation, ResMatching is designed to learn a strong structural prior from fluorescence data rather than to invert a known forward operator. The paper emphasizes four consequences of that design: a learned, structure-aware prior; a superior trade-off between data fidelity and perceptual realism; posterior sampling for diverse plausible HR reconstructions; and calibrated pixel-wise uncertainty estimates. This places the method between deterministic point estimators, which cannot represent uncertainty, and adversarial super-resolution models, which may hallucinate details (Ray et al., 30 Oct 2025).

A common misconception is to treat CSR outputs as direct recoveries of hidden ground truth. The paper explicitly cautions against that interpretation: because high-frequency content is unobserved, visually plausible outputs should be interpreted as informed hypotheses rather than ground truth. This caution is central rather than ancillary, since the method’s posterior sampling and uncertainty maps are motivated by the same irreducible ambiguity.

2. Guided conditional flow matching

ResMatching models the HR conditional distribution through an ODE driven by a conditional velocity field,

$\frac{d x_t}{d t} = v_\theta(t, x_t, x_{M_0}),$

with $x_0 \sim N(0, I)$ and $x_1$ the reconstructed HR image. The interpolation path between a Gaussian base sample and the HR target is defined by

$x_t = (1 - t) x_0 + t x_{M_1}, \qquad t \in [0,1],$

which induces the conditional path distribution

$p_t(x \mid x_{M_1}) = N(t x_{M_1}, (1 - t)^2 I).$

The corresponding conditional flow-matching objective minimizes

$x_{M_1} = H_{M_1}(s) + \eta(s),$0

where $x_{M_1} = H_{M_1}(s) + \eta(s),$1, $x_{M_1} = H_{M_1}(s) + \eta(s),$2, and $x_{M_1} = H_{M_1}(s) + \eta(s),$3 are paired samples (Ray et al., 30 Oct 2025).

The guidance mechanism is implemented directly by conditioning the velocity field on the LR input, $x_{M_1} = H_{M_1}(s) + \eta(s),$4. No classifier-based or classifier-free guidance is used or reported. The degradation $x_{M_1} = H_{M_1}(s) + \eta(s),$5 and the signal-dependent noise $x_{M_1} = H_{M_1}(s) + \eta(s),$6 are unknown and are not explicitly modeled during training; no explicit likelihood $x_{M_1} = H_{M_1}(s) + \eta(s),$7 is assumed or used. This is the key methodological distinction from conditional flow or diffusion formulations that assume known degradations or noise levels. The paper attributes the method’s robustness to noisy LR inputs, especially on the MT-Noisy subset, to this direct conditioning on the observed LR image during transport (Ray et al., 30 Oct 2025).

Training samples $x_{M_1} = H_{M_1}(s) + \eta(s),$8, $x_{M_1} = H_{M_1}(s) + \eta(s),$9, and $H_{M_0}$0, then draws paired $H_{M_0}$1 and $H_{M_0}$2 from the same sample, forms $H_{M_0}$3, and optimizes the velocity-matching loss. Inference initializes $H_{M_0}$4 and performs explicit Euler ODE integration with uniform time steps, returning $H_{M_0}$5 as the CSR prediction. The interpolation time $H_{M_0}$6 is positionally encoded and injected into all residual blocks by adding it to their output tensors; beyond this temporal conditioning, the paper does not provide further architectural specifics (Ray et al., 30 Oct 2025).

3. Data, implementation, and reconstruction protocol

The reported experiments use four BioSR structures: Clathrin-Coated Pits (CCP), Endoplasmic Reticulum (ER), F-actin, and Microtubule-Noisy (MT-Noisy), where MT-Noisy adds additional noise to BioSR’s microtubule data (Ray et al., 30 Oct 2025).

Subset	Raw images (1004×1004)	Patches (128×128)
CCP	39	3120
ER	53	4240
F-actin	35	2800
MT-Noisy	40	3200

Each subset uses 5 validation images and 10 test images. Training uses the flow-matching step-size parameter $H_{M_0}$7 for sampling $H_{M_0}$8. At inference time, full test images are processed with inner tiling at 50% overlap; only the central $H_{M_0}$9 region from each $H_{M_1}$0 tile is used to assemble the final prediction in order to avoid boundary artifacts. Distortion metrics, specifically PSNR and MicroMS-SSIM, are computed on stitched full images, whereas perceptual metrics, LPIPS and FID, are computed on $H_{M_1}$1 inner tiles to avoid tiling seams (Ray et al., 30 Oct 2025).

The reported runtime is $H_{M_1}$2 seconds per full test image on an NVIDIA V100 GPU. Optimizer choice, learning rate, batch size, epochs, memory footprint, and code availability are not specified. This suggests that the paper prioritizes the formulation and empirical behavior of guided conditional flow matching over a detailed engineering description of the network backbone or training system.

4. Evaluation against baselines

ResMatching is compared against seven baselines spanning several CSR paradigms: UNet and RCAN as point predictors; ESRGAN as a generative or adversarial model; INDI with 1-step and 20-step inference as implicit diffusion; Hierarchical VAE (HVAE) as a variational model; and SIFM with $H_{M_1}$3 as flow matching (Ray et al., 30 Oct 2025).

The evaluation uses PSNR and MicroMS-SSIM as fidelity measures, and LPIPS and FID as perceptual measures. Across CCP, ER, F-actin, and MT-Noisy, the reported outcome is consistent: ResMatching achieves competitive fidelity and very strong perceptual quality, yielding the best perception–distortion trade-off in all cases. The paper further notes that the method is particularly effective when the prior is hard to learn, especially when LR images contain substantial noise, as in MT-Noisy. In those cases, it preserves filament continuity and realistic texture while avoiding the over-smoothing associated with UNet and RCAN and the hallucination risk associated with adversarial models (Ray et al., 30 Oct 2025).

Two analytical results qualify these comparisons. First, varying the number of ODE integration steps $H_{M_1}$4 at inference modulates the perception–distortion trade-off, consistent with prior observations for flow and diffusion models. Second, an approximate MMSE estimate obtained by averaging 50 posterior samples improves fidelity metrics over single samples, which the paper identifies as expected for MMSE estimators. The paper does not report separate ablations on architectural components, and it does not report any explicit guidance weights because the only guidance mechanism is conditioning on $H_{M_1}$5.

5. Posterior sampling and calibrated uncertainty

ResMatching is not limited to a single HR prediction. Multiple conditional HR samples are produced by rerunning the ODE integration from different Gaussian initializations $H_{M_1}$6 while conditioning on the same LR image:

$H_{M_1}$7

The paper also reports an approximate MMSE estimate computed as the pixel-wise average of 50 posterior samples (Ray et al., 30 Oct 2025).

Pixel-wise uncertainty is estimated from those posterior samples. For pixel or location $H_{M_1}$8,

$H_{M_1}$9

and

$\eta(s)$0

The analysis then aggregates pixel-wise standard deviations across images into the root mean variance (RMV) and compares RMV to the root mean squared error (RMSE) against HR ground truth. Calibration is assessed by fitting a linear mapping

$\eta(s)$1

The reported reliability plots, comparing RMV and RMSE, are close to the ideal $\eta(s)$2 after calibration, which is presented as evidence that the uncertainty is well calibrated across all tested datasets (Ray et al., 30 Oct 2025).

This makes the uncertainty estimate operational rather than merely descriptive. A plausible implication is that the model’s posterior variance can serve as a rejection signal for uncertain predictions in downstream microscopy workflows, which is also how the paper frames its utility.

6. Position within the literature, limitations, and name ambiguity

ResMatching adopts the conditional flow matching framework and provides a concrete CSR instantiation, but it does not introduce new theoretical guarantees beyond the standard CFM formulation. Its primary contribution is therefore methodological and empirical: a guided conditional flow-matching model for fluorescence-microscopy super-resolution under unknown degradations and heavy noise, together with posterior sampling and calibrated uncertainty (Ray et al., 30 Oct 2025).

Its stated limitations follow directly from the problem formulation. The method relies on paired LR–HR training data. It does not explicitly model the imaging operator $\eta(s)$3 or the noise distribution $\eta(s)$4. CSR remains fundamentally uncertain because high-frequency content is unobserved. Runtime is reported, but memory footprint is not. Code availability, licenses, and additional reproducibility assets are not specified (Ray et al., 30 Oct 2025).

The term “ResMatching” is also ambiguous across recent arXiv literature. It denotes real-time regular expression matching in automata and hardware design (Bernadotte, 2023); it appears as a misspelling of “ReMatching” in scalable shape correspondence (Maggioli et al., 2023); and closely related names, “ReMatch,” designate residual distribution matching for probabilistic downscaling (Kim et al., 29 Jun 2026) and a multimodal retrieval framework based on generative matching in multimodal LLMs (Liu et al., 24 Nov 2025). In fluorescence microscopy, however, ResMatching refers specifically to the CSR method based on guided conditional flow matching (Ray et al., 30 Oct 2025).

A final misconception addressed by the paper is that uncertainty quantification is secondary to reconstruction quality. In this work, posterior sampling, approximate MMSE estimation, and calibrated pixel-wise uncertainty are integral to the method’s definition. They formalize the fact that super-resolution under unknown blur and signal-dependent noise is not merely a single-image restoration problem, but a conditional generative inference problem with irreducible ambiguity.