arxiv: 2601.15844 · v2 · submitted 2026-01-22 · 🌌 astro-ph.IM · astro-ph.CO

Recognition: 2 theorem links

· Lean Theorem

Radio-Interferometric Image Reconstruction with Denoising Diffusion Restoration Models

Michel Morales , Emma Tolley , Remi Poitevineau

Authors on Pith no claims yet

Pith reviewed 2026-05-16 12:21 UTC · model grok-4.3

classification 🌌 astro-ph.IM astro-ph.CO

keywords radio interferometryimage reconstructiondenoising diffusion modelsposterior samplingVLA FIRSTEHTALMACLEAN

0 comments

The pith

A diffusion model trained on VLA radio galaxies reconstructs interferometric images with higher fidelity than CLEAN.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents a data-driven approach to radio interferometric image reconstruction that learns a prior distribution over radio galaxy images using denoising diffusion probabilistic models trained on the VLA FIRST survey. It then applies Denoising Diffusion Restoration Models to perform posterior sampling on simulated incomplete Fourier measurements from VLA, EHT, and ALMA instruments. This yields reconstructions that incorporate the measurement physics directly and achieve higher fidelity than conventional methods operating on gridded visibilities. A reader would care because improved image recovery from sparse radio data could sharpen views of distant sources and compact objects without relying on hand-tuned regularization.

Core claim

The central claim is that a DDPM prior learned from real VLA FIRST radio galaxy observations, when paired with the unsupervised DDRM posterior sampling procedure, reconstructs images from simulated VLA, EHT, and ALMA interferometric data at very high fidelity and with marked improvement over image reconstruction techniques that work on gridded visibilities such as CLEAN.

What carries the argument

Denoising Diffusion Restoration Models (DDRM) that treat the trained DDPM as a learned prior and perform posterior sampling by reversing the diffusion process while respecting the Fourier measurement operator.

If this is right

Reconstructed images achieve very high fidelity on simulated data across multiple radio interferometers.
Performance exceeds that of CLEAN when both are applied to the same gridded visibility data.
The reconstruction remains agnostic to the precise form of the measured visibilities.
Measurement physics enters the sampling process directly through the forward operator.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The method's reliability on real data hinges on whether the FIRST survey prior generalizes to source populations not represented in the training set.
Combining the diffusion prior with multi-frequency or polarization constraints could further constrain reconstructions of complex extended emission.
If the approach scales to full-scale survey data volumes, it could reduce the need for manual CLEAN parameter tuning in routine imaging pipelines.

Load-bearing premise

The distribution of radio galaxy images in the VLA FIRST survey serves as an appropriate prior for the simulated sources observed by VLA, EHT, and ALMA, and the DDRM sampling step correctly embeds the measurement physics without systematic bias.

What would settle it

A test on real (non-simulated) interferometric observations where the DDRM reconstructions show systematic differences from independent reconstructions or from known source structures that cannot be attributed to noise.

Figures

Figures reproduced from arXiv: 2601.15844 by Emma Tolley, Michel Morales, Remi Poitevineau.

**Figure 1.** Figure 1: Neural Network Architecture used for the DDPM network. For components with zero singular value (𝑠𝑖 = 0), we have no information from the measurements y and sample according to the prior only: 𝑝 𝜃 x¯ (𝑖) 𝑡 | x𝑡+1, y = N [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 3.** Figure 3: Diagram of the sampling matrix S, dirty beam F −1S, true imagex, dirty image x𝐷 = F −1 (S) ∗x for the three telescopes considered in this work. 2.6 Metrics We use three metrics for evaluating the quality of the results. The first is Mean Squared Error (MSE), defined as: MSE = 1 𝑁𝑀 ∑︁𝑁 𝑖=1 ∑︁𝑀 𝑗=1 x ( 𝑗) − xˆ ( 𝑗) 𝑖 2 , (16) where x is our true image, xˆ𝑖 is our 𝑖th prediction, and 𝑗 indexes the 𝑀 pixels… view at source ↗

**Figure 4.** Figure 4: DDRM reconstruction results using 1000 sampling steps for four radio galaxies from the test data set using a simulated VLA observation, with no additional noise added. Columns from left to right are: the true image x, the dirty image x𝐷, the DDRM restored image, the residual between the DDRM image and the true image, the mean DDRM image, the residual between the mean DDRM image and the true image, the per-… view at source ↗

**Figure 5.** Figure 5: Reconstruction results for a radio galaxy from the test data set using a simulated VLA observation, with no additional noise added. Columns from left to right are: the true image x, the dirty image x𝐷, the DDRM restored image and its residuals, and then the CLEAN, multi-scale CLEAN, and IUWT compressed sensing reconstructions implemented with the Radler library. The Radler reconstructions are calculated fo… view at source ↗

**Figure 6.** Figure 6: Average MSE per image for 200 reconstructed images of the test data set. Reconstruction is done from a simulated VLA observation with noise from 𝜎𝑦 = 0 to 0.1. The DDRM restored image uses the true value of 𝜎𝑦 (solid blue), or uses 𝜎𝑦 = 0 regardless of the true noise level (diagonal hashed blue). The CLEAN, multi-scale CLEAN, and IUWT compressed sensing reconstructions are calculated for 100,000 iterations… view at source ↗

**Figure 7.** Figure 7: Reconstruction results for a 150 × 150 pixel cutout of 3c353, an out-of-domain radio galaxy, imaged from a simulated VLA observation, variable observation noise added from 𝜎𝑦 = 0 to 0.1. Columns from left to right are: the true image x, the dirty image x𝐷, the DDRM restored image using the true value of 𝜎𝑦, the DDRM restored image using 𝜎𝑦 = 0, and then the CLEAN, multi-scale CLEAN, and IUWT compressed sen… view at source ↗

**Figure 8.** Figure 8: MSE as a function of true image pixel brightness for (left) 200 images of the test data set and (right) several different 150 × 150 pixel cutouts of 3c353 and. Reconstruction is done from a simulated VLA observation with noise from 𝜎𝑦 = 0 to 0.1. The DDRM restored image uses the true value of 𝜎𝑦 (solid blue), or uses 𝜎𝑦 = 0 regardless of the true noise level (blue line). The CLEAN, multi-scale CLEAN, and I… view at source ↗

**Figure 1.** Figure 1: Reconstructing a 150 × 150 pixel image of 3c353 using uniform (all uv bins get the same weight) and natural (uv bins are weighted according to the number of uv samples) weighting, with 𝜎𝑦 = 0 (top) and 𝜎𝑦 = 0.05 (bottom). in [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗

read the original abstract

Reconstructing images of the radio sky from incomplete Fourier information is a key challenge in radio astronomy. In this work, we present a method for radio interferometric image reconstruction using a data-driven prior for the radio sky based on denoising diffusion probabilistic models (DDPMs). We train a DDPM on radio galaxy observations from the VLA FIRST survey, then create simulated VLA, EHT, and ALMA observations of radio galaxies. We use an unsupervised posterior sampling method called Denoising Diffusion Restoration Models (DDRM) to reconstruct the corresponding images using our DDPM as a prior. Our approach is agnostic to the measured radio interferometric data and naturally incorporates the physics of the measurement process. We are able to reconstruct images with very high fidelity and demonstrate a marked improvement over image reconstruction techniques that work on gridded visibilities like CLEAN.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

DDPM trained on FIRST data inside DDRM gives a new data-driven prior for radio interferometry, but missing metrics and frequency mismatch with EHT/ALMA are real problems.

read the letter

The paper's core move is training a DDPM on VLA FIRST survey images and then using DDRM to draw posterior samples for simulated VLA, EHT, and ALMA visibilities. That specific combination looks new in this subfield and folds the known interferometer response directly into the sampling step without retraining the prior each time. Using real survey data for the prior is also cleaner than the usual hand-designed regularizers. Those are the parts that actually advance the work. The abstract's claim of very high fidelity and marked improvement over CLEAN is the part that does not land. No numbers, no error bars, no ablation tables, and no direct comparison metrics appear in the provided text, so it is impossible to judge whether the improvement is real or just visual. The larger concern is the domain gap the stress-test note flags. FIRST is 1.4 GHz at arcsecond scales; EHT is 230 GHz at microarcseconds and ALMA sits in the sub-mm. Radio galaxy structure changes with frequency and resolution through spectral index and optical-depth effects. If the learned manifold does not contain those variants, the sampler will likely insert FIRST-like features rather than recover the true sky. The paper calls the method agnostic to the data, but without cross-frequency validation that claim stays untested. The measurement operator itself follows standard DDRM and the prior comes from independent public data, so there is no obvious circularity or fitting artifact. Still, the central results rest on simulations whose realism relative to real instrumental effects cannot be checked from what is shown. This is worth sending to referees who know both radio interferometry and diffusion models. A serious editor should route it to review rather than desk-reject, but the authors will need to supply quantitative metrics and domain-shift tests before it can be taken as reliable. I would bring the full paper to a reading group only if the figures and tables actually deliver the missing numbers.

Referee Report

3 major / 2 minor

Summary. The paper proposes using a denoising diffusion probabilistic model (DDPM) trained on VLA FIRST survey images as a data-driven prior within the Denoising Diffusion Restoration Models (DDRM) framework for radio-interferometric image reconstruction. It generates simulated VLA, EHT, and ALMA observations of radio galaxies and applies unsupervised DDRM posterior sampling to recover images, claiming very high fidelity and marked improvement over gridded-visibility methods such as CLEAN. The method is presented as agnostic to the specific data while incorporating the interferometer measurement physics.

Significance. If the central claims hold after addressing domain-shift concerns, the work would introduce a promising learned-prior approach to radio imaging that naturally embeds the linear measurement operator, potentially improving reconstruction of complex morphologies where traditional methods struggle. The use of an independent public survey for the prior and the unsupervised sampling step are positive elements that avoid direct data-fitting circularity.

major comments (3)

[Abstract] Abstract: the claims of 'very high fidelity' and 'marked improvement' over CLEAN lack any supporting quantitative metrics (e.g., PSNR, SSIM, normalized mean squared error), error bars, or statistical tests across the simulated datasets; without these, the headline result cannot be evaluated.
[§2 and §3] §2 (DDPM training) and §3 (DDRM application): the prior is learned exclusively from VLA FIRST 1.4 GHz images, yet the method is applied to simulated EHT (230 GHz) and ALMA observations; no cross-frequency validation, ablation on spectral-index effects, or resolution-matching tests are described, raising the risk that reconstructions hallucinate FIRST-like structures rather than recover true sky morphology.
[§4] §4 (Results): visual comparisons alone are insufficient to support superiority over CLEAN; quantitative tables or plots comparing reconstruction fidelity across VLA/EHT/ALMA cases, including controls with mismatched priors, are needed to substantiate the central claim.

minor comments (2)

[Abstract] Abstract: the phrase 'agnostic to the measured radio interferometric data' is imprecise; DDRM explicitly uses the known measurement operator, so the wording should be clarified to avoid confusion.
Figure captions and text should explicitly state the number of independent simulations, noise realizations, and any hyperparameter choices for the DDPM noise schedule to allow reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their thoughtful and constructive comments on our manuscript. We address each of the major comments below and will make revisions to incorporate quantitative evaluations and additional validations as suggested.

read point-by-point responses

Referee: [Abstract] Abstract: the claims of 'very high fidelity' and 'marked improvement' over CLEAN lack any supporting quantitative metrics (e.g., PSNR, SSIM, normalized mean squared error), error bars, or statistical tests across the simulated datasets; without these, the headline result cannot be evaluated.

Authors: We agree with this assessment. The current manuscript relies primarily on visual comparisons for the claims in the abstract. In the revised version, we will add quantitative metrics including PSNR, SSIM, and normalized mean squared error, along with error bars from multiple simulation runs and statistical comparisons to CLEAN. This will provide the necessary support for the headline results. revision: yes
Referee: [§2 and §3] §2 (DDPM training) and §3 (DDRM application): the prior is learned exclusively from VLA FIRST 1.4 GHz images, yet the method is applied to simulated EHT (230 GHz) and ALMA observations; no cross-frequency validation, ablation on spectral-index effects, or resolution-matching tests are described, raising the risk that reconstructions hallucinate FIRST-like structures rather than recover true sky morphology.

Authors: This point highlights an important limitation in the current presentation. Although radio galaxy morphologies are expected to be similar across frequencies for the structures we consider, we did not include explicit cross-frequency tests. We will revise the manuscript to include a discussion of potential domain shifts, an ablation study varying spectral indices in the simulations, and resolution-matching experiments to confirm that the reconstructions recover true morphology rather than imposing FIRST-like features. revision: yes
Referee: [§4] §4 (Results): visual comparisons alone are insufficient to support superiority over CLEAN; quantitative tables or plots comparing reconstruction fidelity across VLA/EHT/ALMA cases, including controls with mismatched priors, are needed to substantiate the central claim.

Authors: We concur that additional quantitative evidence is required. We will expand Section 4 to include tables and plots with fidelity metrics for all three instruments (VLA, EHT, ALMA). Furthermore, we will add control experiments using mismatched priors (e.g., a diffusion model trained on non-radio data) to demonstrate that the performance gains stem from the radio-specific prior learned from the FIRST survey. revision: yes

Circularity Check

0 steps flagged

No significant circularity; prior from independent survey and measurement physics applied separately

full rationale

The paper trains a DDPM on the external VLA FIRST survey to learn a data-driven prior, then applies DDRM posterior sampling that uses the known linear measurement operator for the interferometer response. No equation or step reduces the reconstructed image to a quantity defined or fitted by the same target data by construction. The prior distribution is independent of the test simulations, and the physics incorporation is standard and external to the learned model, making the overall derivation self-contained.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the learned image distribution serving as a valid prior and on the correctness of the DDRM sampling procedure for the radio measurement operator.

free parameters (1)

DDPM noise schedule and network weights
The diffusion model parameters are fitted to the FIRST survey images and directly shape the prior used in reconstruction.

axioms (1)

domain assumption Radio sky images are drawn from the same distribution as the VLA FIRST survey galaxies
Invoked when the trained DDPM is applied to simulated observations from other arrays.

pith-pipeline@v0.9.0 · 5442 in / 1283 out tokens · 34491 ms · 2026-05-16T12:21:59.212583+00:00 · methodology

discussion (0)

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

AstroSURE: Learning to Remove Noise from Astronomical Images Without Ground Truth Data
astro-ph.IM 2026-04 unverdicted novelty 5.0

Unsupervised denoising methods improve faint-source detection in astronomical images from HST and CFHT, with better performance when models are initialized on similar-domain data.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages · cited by 1 Pith paper · 5 internal anchors

[1]

Aghabiglou,A.,Chu,C.S.,Dabbech,A.,&Wiaux,Y.,2024.TheR2D2Deep Neural Network Series Paradigm for Fast Precision Imaging in Radio Astronomy,The Astrophysical Journal Supplement Series,273(1),

work page 2024
[2]

Becker, R

Bagchi,S.&Mitra,S.K.,1999.TheNonuniformDiscreteFourierTransform and Its Applications in Signal Processing, The Springer International Series in Engineering and Computer Science, Springer, New York, NY, 1st edn. Becker, R. H., White, R. L., & Helfand, D. J.,

work page 1999
[3]

Matchingpursuitofimages,inProceedings., International Conference on Image Processing, vol

Bergeaud,F.&Mallat,S.,1995. Matchingpursuitofimages,inProceedings., International Conference on Image Processing, vol. 1, pp. 53–56 vol.1. Bhatnagar, S., Cornwell, T. J., Golap, K., & Uson, J. M.,

work page 1995
[4]

Sparsity averaging reweighted analysis (sara): a novel algorithm for radio-interferometric imaging: Sara for radio-interferometric imaging,Monthly Notices of the Royal Astronomical Society,426(2), 1223–1234. Chung,H.,Kim,J.,Mccann,M.T.,Klasky,M.L.,&Ye,J.C.,2023.Diffusion posterior sampling for general noisy inverse problems, inThe Eleventh International C...

work page 2023
[5]

Cornwell, T

Deep radio- interferometric imaging with POLISH: DSA-2000 and weak lensing, MNRAS,514(2), 2614–2626. Cornwell, T. J.,

work page 2000
[6]

Cornwell,T.J.&Perley,R.A.,1992

A simple maximum entropy deconvo- lution algorithm,A&A,143(1), 77–83. Cornwell,T.J.&Perley,R.A.,1992. Radio-interferometricimagingofvery large fields. The problem of non-coplanar arrays.,A&A,261, 353–364. Cornwell,T.J.,Golap,K.,&Bhatnagar,S.,2008. Thenoncoplanarbaselines effectinradiointerferometry:Thew-projectionalgorithm,IEEEJournal of Selected Topics i...

work page 1992
[7]

Deepresiduallearningforimage recognition, in2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

He,K.,Zhang,X.,Ren,S.,&Sun,J.,2016. Deepresiduallearningforimage recognition, in2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778. Ho,J.,Jain,A.,&Abbeel,P.,2020. Denoisingdiffusionprobabilisticmodels. Högbom, J. A.,

work page 2016
[8]

Lei Ba, J., Kiros, J

Kawar,B.,Elad,M.,Ermon,S.,&Song,J.,2022.Denoisingdiffusionrestora- tion models, inAdvances in Neural Information Processing Systems. Lei Ba, J., Kiros, J. R., & Hinton, G. E.,

work page 2022
[9]

Layer Normalization

Layer Normalization,arXiv e-prints, p. arXiv:1607.06450. Li, F., Cornwell, T. J., & de Hoog, F.,

work page internal anchor Pith review Pith/arXiv arXiv
[10]

SGDR: Stochastic Gradient Descent with Warm Restarts

SGDR: Stochastic Gradient Descent with Warm Restarts,arXiv e-prints, p. arXiv:1608.03983. Loshchilov, I. & Hutter, F.,

work page internal anchor Pith review Pith/arXiv arXiv
[11]

Decoupled Weight Decay Regularization

Decoupled Weight Decay Regularization, RASTI000, 1–10 (2026) Radio-Interferometric Image Reconstruction with Denoising Diffusion Restoration Models9 arXiv e-prints, p. arXiv:1711.05101. Narayan,R.&Nityananda,R.,1986. Maximumentropyimagerestorationin astronomy.,ARA&A,24, 127–170. Offringa, A. R. & Smirnov, O.,

work page internal anchor Pith review Pith/arXiv arXiv 2026
[12]

FiLM: Visual Reasoning with a General Conditioning Layer

Pang,T.,Lu,C.,Du,C.,Lin,M.,YAN,S.,&Deng,Z.,2023. Oncalibrating diffusion probabilistic models, inThirty-seventh Conference on Neural Information Processing Systems. Perez,E.,Strub,F.,deVries,H.,Dumoulin,V.,&Courville,A.,2017. FiLM: Visual Reasoning with a General Conditioning Layer,arXiv e-prints, p. arXiv:1709.07871. Ponsonby,J.E.B.,1973. Anentropymeasur...

work page internal anchor Pith review Pith/arXiv arXiv 2023
[13]

arXiv:2601.07485

A Guided Unconditional Diffusion Model to Synthesize and Inpaint Radio Galaxies from FIRST, MGCLS and Radio Zoo,arXiv e-prints, p. arXiv:2601.07485. Rich,J.W.,deBlok,W.J.G.,Cornwell,T.J.,Brinks,E.,Walter,F.,Bagetakos, I., & Kennicutt, Jr., R. C.,

work page arXiv
[14]

U-Net: Convolutional Networks for Biomedical Image Segmentation

U-Net: Convolutional Networks for Biomedical Image Segmentation,arXiv e-prints, p. arXiv:1505.04597. Schmidt, K., Geyer, F., Fröse, S., Blomenkamp, P.-S., Brüggen, M., de Gasperin, F., Elsässer, D., & Rhode, W.,

work page internal anchor Pith review Pith/arXiv arXiv
[15]

The Undecimated Wavelet De- composition and its Reconstruction,IEEE Transactions on Image Pro- cessing,16(2), 297–309. Taran, O., Bait, O., Dessauges-Zavadsky, M., Holotyak, T., Schaerer, D., & Voloshynovskiy,S.,2023.Challenginginterferometricimaging:Machine learning-based source localization from uv-plane observations,A&A, 674, A161. Teneggi,J.,Tivnan,M....

work page arXiv 2023
[16]

VičánekMartínez,T.,Edler,H.W.,&Brüggen,M.,2025.Simulatingrealistic radio continuum survey maps with diffusion models,A&A,700, A18

Simulating images of radio galaxies with diffusion models,Astronomy and Astro- physics,691, A360. VičánekMartínez,T.,Edler,H.W.,&Brüggen,M.,2025.Simulatingrealistic radio continuum survey maps with diffusion models,A&A,700, A18. Wang,R.,Chen,Z.,Luo,Q.,&Wang,F.,2023.Aconditionaldenoisingdiffu- sion probabilistic model for radio interferometric image recons...

work page 2025
[17]

APPENDIX A: UNIFORM VS NATURAL UV WEIGHTS Thetransformationoperatorconsideredinthroughoutthepapercor- responds touniformweighting, giving all uv bins the same weight

Regularized Maxi- mum Likelihood Image Synthesis and Validation for ALMA Continuum Observations of Protoplanetary Disks,PASP,135(1048), 064503. APPENDIX A: UNIFORM VS NATURAL UV WEIGHTS Thetransformationoperatorconsideredinthroughoutthepapercor- responds touniformweighting, giving all uv bins the same weight. Different weight schemes are easily implemente...

work page 2026