Pixellated Posterior Sampling of Point Spread Functions in Astronomical Images
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Astrophysics. The day in astrophysics: observations, models and what the data forces.
Vera: We're looking at the paper: "Pixellated Posterior Sampling of Point Spread Functions in Astronomical Images".
Jocelyn: Here's the gist of the paper. We introduce a novel framework for upsampled Point Spread Function (PSF) modeling using pixellevel Bayesian inference, which enables the first meaningful propagation of detailed PSF morphological uncertainty in downstream analysis Pruning redundant reasoning tokens Accurate characterization of the PSF remains a persistent challenge in observational astronomy The effective PSF (ePSF) exhibits significant spatiotemporal variability driven by mechanical and thermal variations over time, such as orbital focus changes or “breathing” This variability means that every observation is convolved with an essentially unique PSF, necessitating observation-specific modeling The core challenge in Bayesian PSF inference is the specification of a realistic prior, P(M) How the training signal is built The authors replace simple priors with a highly expressive, pixel-level prior P(M) learned from high resolution sample images They leverage score-based generative diffusion models to learn this complex morphological feature, which implicitly represents this complex prior through a learned “score function”, ∇M log(P(M)) A key requirement for this approach is a representative training set, using the 9,776 individual ePSF templates described in Section 4.2 The diffusion model is trained to learn only the deviations from a mean template M¯, which is computed as an average ePSF template sample Training duration proved to be a critical hyperparameter, and extended training caused the model to overfit Where the gains come from The method achieves orders of magnitude higher likelihood and residuals consistent with noise while remaining visually realistic Traditional methods, such as simple parametric fitting (Moffat), fixed template libraries (ePSF), and empirical regularized likelihood methods (PSFex), show a consistent ∼ 5% flux bias The posterior mean flux provides an accurate, unbiased flux measurement with a principled uncertainty estimate which propagates morphological uncertainty Quantifying morphological uncertainty The framework produces a statistically sufficient model that fully explains the data, as shown by residual images exhibiting essentially no discernible structure and χ2 / dof is approximately one for all examples The method yields a principled, pixel-level quantification of morphological uncertainty which may be marginalized as in column 4 A non-parametric, weighted mean scheme is adopted for interpolating the PSF posterior to arbitrary detector locations This weighted interpolation scheme is superior even to the fitting results of traditional models, with residuals that are far more consistent with noise The availability of full, quantified PSF posterior distributions holds transformative implications across a number of astronomical disciplines Pruning redundant reasoning tokens The authors successfully demonstrated a novel framework that produces pixel-level Bayesian posterior Point Spread Functions (PSF) in HST imagery using a Gaussian likelihood and generative diffusion model prior This approach was benchmarked against three traditional PSF modeling techniques: simple parametric fitting (Moffat), fixed template libraries (ePSF), and empirical regularized likelihood methods (PSFex) The production of a full posterior distribution for the PSF morphology enables the previously intractable marginalization over model uncertainty in subsequent analyses This is not merely a formal improvement; we demonstrate in Appendix A that traditional methods can introduce systematic flux biases of order 5% Our procedure also yields unbiased measurements for significantly fainter stars than traditional methods, simply producing a broader posterior in accordance with the lower signal-to-noise The primary limitation identified in this work is the use of a non-parametric, ad-hoc weighted mean for interpolating the PSF posterior to arbitrary image locations A leave-one-out cross-validation test in Appendix B demonstrates this interpolation scheme is superior even to the fitting results of traditional models The most formal approach would be a PSF prior conditioned on detector position; given a likelihood at the position of individual stars, this would produce posterior realizations of the full frame variable PSF model The availability of full, quantified PSF posterior distributions holds transformative implications across a number of astronomical disciplines Pruning redundant reasoning tokens
Title and authors: Vera: So, Jocelyn, this paper is called "Pixellated Posterior Sampling of Point Spread Functions in Astronomical Images," and it’s about using a new framework for modeling those messy point spread functions—the shape of light as it gets through a telescope and an atmosphere.
Jocelyn: Right. It’s essentially proposing a way to use pixel-level Bayesian inference to get the posterior distribution of the PSF, which is how uncertain we are about that shape, in astronomical images. It’s tackling that big problem where every single observation has its own unique PSF because of things like orbital changes or thermal effects.
Subrahmanyan: That variability means we can't just use one fixed template anymore; we need a model that accounts for the specific conditions of each exposure, which is exactly what this paper seems to be aiming at by defining a pixel-level prior P(M) <ref:2511.19594#pg1>.
Vera: It’s interesting how they are using score-based generative diffusion models to learn that complex morphological prior, instead of just using simple priors like the Moffat function or fixed libraries like ePSF templates <ref:2511.19594#pg2>.
Jocelyn: They train the diffusion model to learn deviations from a mean template, M bar, which is basically an average PSF sample, and they do this using nine thousand seven hundred seventy-six individual ePSF templates as training data <ref:2511.19594#pg2>.
Subrahmanyan: The authors say that this approach allows them to get orders of magnitude higher likelihoods and residuals that are consistent with noise, which is a big deal for achieving high fidelity reconstruction <ref:2511.19594#pg3>.
Vera: And they show that traditional methods like the Moffat model or the PSFex method introduce a systematic flux bias of about five percent, so this new approach provides an unbiased flux measurement with a principled uncertainty estimate <ref:2511.19594#pg3>.
Jocelyn: It also produces pixel-level quantification of morphological uncertainty, which they say can be marginalized later, and they use a non-parametric, weighted mean scheme to interpolate that posterior to arbitrary detector locations <ref:2511.19594#pg3>.
Subrahmanyan: That full posterior distribution is what really matters for structure analysis; it enables the marginalization over model uncertainty in subsequent analyses, which is crucial for things like finding dark matter subhalos in strong gravitational lensing <ref:2511.19594#pg2>.
Vera: But they admit a limitation: they use a non-parametric, ad hoc weighted mean for that interpolation to arbitrary locations, and while leave-one-out cross-validation shows it’s better than traditional fitting results, the authors suggest that conditioning the prior on detector position would be more formal <ref:2511.19594#pg3>.
Jocelyn: It sounds like a lot of computational heavy lifting to build and train that generative model, but the result is a PSF posterior that has residuals looking almost exactly like noise when compared to traditional methods <ref:2511.19594#pg3>.
Subrahmanyan: The implication for cosmology or structure formation is that we can start making precision shape measurements in weak lensing surveys because the PSF uncertainty, which used to be a major source of systematic error, is now properly propagated <ref:2511.19594#pg2>.
Vera: So, to wrap up this paper on "Pixellated Posterior Sampling of Point Spread Functions in Astronomical Images," they’ve shown how pixel-level Bayesian inference with a diffusion model prior can yield unbiased flux measurements and a principled way to handle morphological uncertainty in PSF modeling <ref:2511.19594#pg3>.
Jocelyn: It opens the door for high-fidelity image reconstruction, allowing us to analyze structures like dark matter subhalos in strong lensing with much more confidence than before <ref:2511.19594#pg2>.
Subrahmanyan: And for me, the main thing is that this gives us a robust tool to get cleaner data on the cosmic web by properly accounting for these instrumental artifacts, provided we can handle that interpolation scheme they used <ref:2511.19594#pg3>.
The paper's summary: Vera: So, we've talked about how this paper uses a pixel level Bayesian inference to model point spread functions in images like those from the Hubble Space Telescope and what that means for things like weak lensing surveys.
Jocelyn: Right. Basically, they’re taking all those messy variations in the PSF caused by things like the telescope moving or heating up and they're creating a statistical picture of what that shape is actually doing at every single pixel in the image.
Subrahmanyan: From a theoretical side, what this really does is it lets us go beyond just knowing *an* average PSF. It gives us the full distribution, the uncertainty about that shape itself, which we can then use to make much more precise measurements of things like galaxy shapes or dark matter distributions.
Vera: And they show that this method produces residuals that look really clean, almost noise-level clean when compared to older methods like fitting with Moffat functions or fixed libraries. That is a big deal for image reconstruction.
Jocelyn: It means we can generate visually realistic models because the math actually matches what the data looks like on a pixel by pixel, which is what you want for things like studying dark matter subhalos in strong gravitational lensing.
Subrahmanyan: But they did have to use a specific way of picking that uncertainty—a non-parametric weighted mean—to place the PSF model onto different locations, and they flag that as something they need to be careful about when applying it broadly.
Vera: That's fair, you can't always have a perfect interpolation scheme across the whole sky without some compromises. But for a single source analysis, they claim this pixel level information gives us an unbiased flux measurement where older techniques systematically introduced about five percent of bias.
Jocelyn: So if we are trying to measure how bright a star is, or how much light is coming from a galaxy, this approach gives us the actual number and tells us exactly how much error there is on that number because it's tied directly to the shape uncertainty.
Subrahmanyan: And that direct link between shape uncertainty and flux measurement is what makes this framework useful for cosmology. If we can accurately model the PSF, we can then subtract its effect more reliably when we try to measure the faint signals from distant objects or weak lensing shear.
Vera: So, in short, they’ve shown a way to get a pixel level understanding of PSF uncertainty that leads to unbiased flux measurements and better shape modeling for things like dark matter studies.
Jocelyn: It’s moving us toward more robust analysis pipelines where the instrumental effects aren't just ignored or crudely corrected anymore.
Subrahmanyan: And that kind of detailed morphological information is exactly what we need if we want to really start using weak lensing surveys for precision shape measurements.
The paper's improvements: Vera: So we’ve seen that this paper introduces a method to get pixel level Bayesian posterior point spread functions in images using diffusion models, which helps us understand the shape uncertainty better than before.
Jocelyn: Exactly. But beyond just getting a better posterior distribution for the PSF itself, the paper suggests some ways to actually use that information more effectively in real analysis pipelines.
Subrahmanyan: I think one of the big improvements they hint at is how we can move from just modeling point sources to modeling entire fields where conditions are changing.
Vera: Yeah, they talk about this being useful for precision shape measurements in weak lensing surveys, which is huge because those measurements are fundamentally limited by how well you know the PSF.
Jocelyn: It means instead of applying a single correction across an image, we can use the pixel level data to correct things locally where the PSF is actually different.
Subrahmanyan: That local correction ability is what unlocks precision shape measurements, which means we can get much tighter constraints on dark matter subhalos in strong gravitational lensing.
Vera: And they also touch on how this helps with robust point source characterization, especially for high contrast imaging or when you’re trying to deblend multiple sources near each other.
Jocelyn: It’s about having a full posterior distribution for every single source you measure, not just a single best guess for its shape.
Subrahmanyan: So it moves us toward making the entire analysis pipeline more principled, rather than relying on empirical fitting methods that can introduce systematic biases we talked about earlier.
Vera: And they also point out that the next step is formalizing the prior—making the PSF prior dependent on where you are in the image, not just being a general model.
Jocelyn: That sounds like they’re looking at conditioning that diffusion model directly on detector position, which would be even more powerful for large-scale surveys.
Subrahmanyan: If they can manage that formally, it means we can get a really principled way to propagate the morphological uncertainty through all the downstream science calculations.
Conclusion: Tom: So we’ve covered how this paper uses pixel level Bayesian inference to model point spread functions in images using diffusion models, which helps us understand the shape uncertainty better than before and gives us more robust flux measurements.
Vera: Exactly. It shows that traditional methods can introduce a five percent bias, so this new approach gives us unbiased numbers while still providing a principled way to estimate the error bars on those measurements.
Jocelyn: And that’s really important for follow-up work because if our initial measurement is biased, the whole chain of discovery downstream gets skewed.
Subrahmanyan: From my side, it means we can start making more precision shape measurements in weak lensing surveys because the PSF uncertainty is no longer a huge systematic error source.
Vera: And they’ve laid out how this can be used to recover pixel level posterior distributions for point sources, which is vital for high contrast imaging and AGN studies.
Jocelyn: It gives us a complete picture of the source shape uncertainty at every pixel, not just one average shape estimate.
Subrahmanyan: That detailed information allows for much better propagation of morphological uncertainty when we try to reconstruct structures like dark matter subhalos in strong lensing.
Vera: They also noted that the main caveat is using a non-parametric weighted mean to interpolate the posterior to arbitrary detector locations, so you have to be careful how you apply it across the whole image.
Jocelyn: It’s a practical limitation they acknowledged, but they showed that even with that interpolation, their results are much cleaner than what traditional fitting methods give us for residuals.
Subrahmanyan: So it’s a powerful tool for cosmology and structure formation because we get a more principled way to handle instrumental effects when we try to map the universe.
Vera: And overall, this paper on Pixellated Posterior Sampling of Point Spread Functions in Astronomical Images gives us a much more reliable way to deal with these complex instrumental artifacts in our data.
Jocelyn: It sets a new standard for how we should be thinking about modeling those messy optical effects when we look at the sky.
Subrahmanyan: We’ll need to see if this approach scales up well for the massive datasets coming from next generation telescopes to really confirm its utility.
Connor Stone, Ronan Legin, Alexandre Adam, Nikolay Malkin, Gabriel Missael Barco, Laurence Perreaul-Levasseur and Yashar Hezaveh
Montreal Institute for Astrophysical Data Analysis and Machine Learning · University of Montreal · University of Toronto · University of Edinburgh · Center for Computational Astrophysics, Flatiron Institute · Trottier Space Institute · Perimeter Institute
astro-ph.IM
Submitted: 2025-11-24
Updated: 2026-11-02
Comments: 21 pages, 12 figures, Published in ApJ
Code: https://github.com/AlexandreAdam/score
License: http://creativecommons.org/licenses/by-nc-sa/4.0/
Importance score: 91/100
The gist: We introduce a novel framework for upsampled Point Spread Function (PSF) modeling using pixellevel Bayesian inference Accurate PSF characterization is critical for precision measurements in many…
Key concepts
- Pixel-level Bayesian Inference
- This technique treats every single pixel in an image as a variable and uses Bayesian statistics to estimate its probability distribution. Instead of fitting a single, averaged model, it generates a full posterior distribution for the PSF morphology at each location.
- Score-based Generative Diffusion Models
- These are advanced AI models trained to understand complex data distributions. In this context, they learn the intricate morphological features of realistic Point Spread Functions by learning a 'score function' that guides the generation of these detailed, spatially varying models.
- Morphological Uncertainty Propagation
- This is the ability to pass information about how uncertain the PSF shape is from one stage of analysis (like image reconstruction) to another. The framework provides a principled, pixel-level quantification of this uncertainty, which traditional methods lack.
- ePSF Variability
- The effective PSF changes over time and space due to mechanical and thermal variations in the telescope. This method is designed to capture this spatiotemporal variability by training on a large library of individual ePSF templates rather than a single average model.
Terminology
Summary
We introduce a novel framework for upsampled Point Spread Function (PSF) modeling using pixellevel Bayesian inference Accurate PSF characterization is critical for precision measurements in many fields including: weak lensing, astrometry, and photometry Our method defines the posterior distribution of the pixelized PSF model through the combination of an analytic Gaussian likelihood and a highly expressive generative diffusion model prior, trained on a library of HST ePSF templates Compared to traditional methods (parametric Moffat, ePSF template-based, and regularized likelihood), we demonstrate that our PSF models achieve orders of magnitude higher likelihood and residuals consistent with noise, all while remaining visually realistic By recovering a realistic, pixel-level posterior distribution, our technique enables the first meaningful propagation of detailed PSF morphological uncertainty in downstream analysis
Introduction to the Problem
All astronomical observations are subject to image degradation arising from the redistribution of true sky flux due to the instrument’s physical characteristics The composite effect of these distortions is formalized as the Point Spread Function (PSF): the observed image of an ideal point-like light source Accurate characterization of the PSF remains a persistent challenge in observational astronomy The effective PSF (ePSF) encompasses not only the optical path PSF but also detector-level effects such as pixel response and charge diffusion For space-based observatories like the Hubble Space Telescope (HST), the ePSF exhibits significant spatiotemporal variability driven by mechanical and thermal variations over time, such as orbital focus changes or “breathing” This variability, which also depends on position in the detector and wavelength, means that every observation is convolved with an essentially unique PSF, necessitating observation-specific modeling The core challenge in Bayesian PSF inference is the specification of a realistic prior, P(M) We replace the simple parametric, regularized, or implicit priors with a highly expressive, pixel-level prior P(M) learned from high resolution sample images This approach enables, for the first time, robust sampling of the full, pixel-level posterior distribution P(MD), yielding substantially improved PSF reconstructions and a principled propagation of morphological PSF uncertainty to downstream analysis
Traditional Modeling Limitations
The traditional methods suffer from severe model misspecification, leading to biased flux measurements The simple Moffat profile leaves pronounced systematic residuals, which frequently exceed 5σ significance for the brightest sources The formal uncertainties computed by AstroPhot for the Moffat parameters are exceedingly small, which is a direct consequence of severe model misspecification Fixed ePSF templates provide only a point estimate of the PSF structure, failing to propagate uncertainties in the PSF’s morphology Empirical regularization methods like PSFex suffer from mixed failure, with some stars having implausibly large residuals and others overfitting the data due to a misspecified prior
How the training signal is built
To define our prior, we leverage score-based generative diffusion models (Y. Song et al. p. 1 To define our prior,
we leverage score-based generative diffusion models (Y. Song et al. Song et al. 2021), which implicitly represent this complex prior through a learned “score function”, ∇M log(P(M)). First, we
apply extensive data augmentation to the training samples, including shifts, flips, scaling, rotations, skews, and
blurring, to expand the effective prior space (each applying to 50% of the training samples). Second, as we will
discuss, we intentionally stop the model’s training early
to prevent it from “memorizing” the discrete templates. Our diffusion model, implemented using a U-Net
architecture (O. Ronneberger et al.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Precise PSF modeling for weak lensing surveys: The method enables
the first meaningful propagation of detailed PSF morphological uncertainty in downstream analysis,
allowing forprecision shape measurements
that are fundamentally limited by PSF knowledge. -
Unbiased flux measurements for faint sources: The Bayesian framework produces fluxes that are
accurate, unbiased measurements
compared to traditional methods, which show asystematic ~ 5% bias.
This is crucial for accurate photometry and avoids propagating significant systematic biases into scientific results. -
High-fidelity image reconstruction: The system can generate
visually realistic models
with residualsconsistent with noise,
enabling the analysis of structures like dark matter subhalos in strong gravitational lensing by providing aprincipled propagation of morphological PSF uncertainty.
-
Robust point source characterization: The technique allows for the recovery of a
pixel-level posterior distribution
for point sources, which is vital for high-contrast imaging and AGN studies requiring precise deblending. -
General applicability across varying instrumental states: By training on a library of HST ePSF templates and using diffusion models, the system produces PSF models that are robust to
spatiotemporal variability driven by mechanical and thermal variations over time.
Abstract
We introduce a novel framework for upsampled Point Spread Function (PSF) modeling using pixel-level Bayesian inference. Accurate PSF characterization is critical for precision measurements in many fields including: weak lensing, astrometry, and photometry. Our method defines the posterior distribution of the pixelized PSF model through the combination of an analytic Gaussian likelihood and a highly expressive generative diffusion model prior, trained on a library of HST ePSF templates. Compared to traditional methods (parametric Moffat, ePSF template-based, and regularized likelihood), we demonstrate that our PSF models achieve orders of magnitude higher likelihood and residuals consistent with noise, all while remaining visually realistic. Further, the method applies even for faint and heavily masked point sources, merely producing a broader posterior. By recovering a realistic, pixel-level posterior distribution, our technique enables the first meaningful propagation of detailed PSF morphological uncertainty in downstream analysis. An implementation of our posterior sampling procedure is available on GitHub.
Sources
- Posterior samples of source galaxies in strong gravitational lenses with score-based priors
- The spatially-resolved effect of mergers on the stellar mass assembly of MaNGA galaxies
- Inpainting Galaxy Counts onto N-Body Simulations over Multiple Cosmologies and Astrophysics
- IRIS: A Bayesian Approach for Image Reconstruction in Radio Interferometry with expressive Score-Based priors
- Score-Based Diffusion Models as Principled Priors for Inverse Imaging
- Strong-Lensing Source Reconstruction with Denoising Diffusion Restoration Models
- Can denoising diffusion probabilistic models generate realistic astrophysical fields?
- U-Net: Convolutional Networks for Biomedical Image Segmentation
- A Data-Driven Prism: Multi-View Source Separation with Diffusion Model Priors
- Principled Probabilistic Imaging using Diffusion Models as Plug-and-Play Priors
Related papers
- On-sky demonstration of a vector Zernike wavefront sensor in a cascaded adaptive optics system
- Euclid Data Release 1 (DR1): Calibration strategy and observations
- The General Antiparticle Spectrometer (GAPS) Antarctic Balloon Payload
- Detection and excision of radio-frequency interference in radio arrays
- Unified matched-filter and excess-power pipeline for gravitational-wave detector characterization
- Ushering in sub-arcsecond angular resolution observatories in the far-infrared: The case for a double-Fourier interferometer