Generating radio continuum survey maps of arbitrary size with latent diffusion models

arXiv:2609.06549 · astro-ph.IM · Submitted 2026-09-06 · Read on arXiv

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: "Generating radio continuum survey maps of arbitrary size with latent diffusion models".

Jocelyn: Here's the gist of the paper. Generating radio continuum survey maps of arbitrary size with latent diffusion models implements a latent diffusion model (LDM) to synthesize realistic radio survey map cutouts and a technique called Sliding Window Iterative (SWIIT) sampling to sample these maps of arbitrary size The method addresses the limitations of fixed-size generative models by enabling the generation of large-scale synthetic radio maps with control over source locations and properties Model Architecture and Training The LDM is composed of a vector-quantized variational autoencoder (VQ-VAE) and a diffusion model operating on the resulting latent representations The VQ-VAE uses an encoder to project images into a smaller latent space, which is then processed by the diffusion model The VQ-VAE architecture features three resolution levels and a latent dimension of 3 channels Training the VQ-VAE involves several loss terms, including reconstruction loss, discriminator loss, and commitment loss The diffusion model denoiser network is implemented as a U-Net that incorporates conditional inputs such as the noise level and the inpainting context The training schedule involves several stages, including an initial run for 2 × 10 5 iterations and a second stage with gradient accumulation for another 2 × 10 5 iterations Sampling Technique The SWIIT-sampling technique is developed to break down the process of sampling large sky survey maps into several individual steps This approach works by sampling individual patches within a sliding window moved with a stride of half the image size, passing overlapping parts as additional input for consistency The procedure involves sequentially sampling the entire image, then moving the window to sample only the right half, and so on until the desired map width is reached The final latent map is decoded into image space using a sliding-window convolutional fashion with a stride of half the latent image size Performance and Control The results show that realistic cutouts of 512 pixel side length can be generated with precise control over source locations and brightness Sampling of arbitrary-sized maps is made possible, showing realistic results with seamless combination of individual samples limited only by occasional sampling artifacts The LDM-sampled images replicate highly realistic source morphologies and associated imaging artifacts Control over source location and brightness properties is facilitated by engineering a catalog context vector that encodes position, flux, and size information Limitations and Future Directions A limitation of the present work is that since the model is trained on observed radio survey maps, there is no explicit ground-truth sky model against which the generated images can be compared The repeating-pattern artifacts observed in some sequentially generated examples may indicate a limitation of the proposed sampling strategy Furthermore, the method is specific to LOFAR data, and application to surveys from other telescopes would require training on corresponding data from those instruments The authors suggest that including minor axis and positional angle for each source could provide a more closely related input signal for better control over source size Data Preparation The training data is obtained from observations made with the Low Frequency Array (LOFAR) radio telescope, specifically from the third data release of the LOFAR Two-Metre Sky Survey (LoTSS DR3) Individual cutouts are extracted from these mosaics, and pixel values are scaled using a function γ to transform the asymmetric distribution into one close to zero-mean and unit variance Information on source positions, fluxes, and sizes is encoded into a catalog context vector which is passed as conditioning input to the denoiser network The training set consists of 152,098 cutouts of 1024 px side length In this work, we implemented a generative model for synthesizing realistic individual radio survey map images of 512 px side length, a size that is to our knowledge unprecedented for generative models in the context of astronomical image generation Further, we developed SWIIT-sampling, a novel technique for sampling radio sky survey maps of arbitrary size through iterative use of an LDM trained for image inpainting By designing a custom function for scaling the pixel values of training images, the large dynamic range of radio maps can be dealt with in a consistent way over the entire dataset This makes radio observational data suitable for computer vision tasks while still conserving universal flux information By engineering a context vector that encodes spatial information of radio sources present on the images, as well as their brightness and size properties, we facilitate effective control over sources present on the sampled images By also including information of sources in the immediate vicinity outside of the FOV, we account for a correct modeling of spillover artifacts and thereby allow for coherent iterative sampling in a sliding-window approach This makes it possible to sample realistic images at arbitrary sizes, circumventing current size limitations of deep learning models for image generation The code implemented and used for this work is publicly available under https://github.com/tmartinezML/glori The authors acknowledge support by the Deutsche Forschungsgemeinschaft under Germany’s Excellence Strategy – EXC 2121 Quantum Universe – 390833306 and via the KISS consortium (05D23GU4) funded by the German Federal Ministry of Education and Research BMBF in the ErUM-Data action plan The paper concludes that while results show remarkable realism and quality, the proposed method works at the expense of high computational cost at both training and inference time The authors also note that consistency models are a promising alternative to alleviate the computational cost associated with sequential denoising procedures The paper is applicable to generating any kinds of large-scale mock image data, given the availability of enough training images The authors also note that the quality of the source catalogs used for training is a limiting factor in this regard The paper provides a method that can be employed as an emulator for otherwise more computationally expensive modeling steps Pruning redundant reasoning tokens The authors describe the LDM as combining DMs with VAEs, where the DM operates inside the latent space of a VQ-VAE The VQ-VAE is responsible for reconstructing fine-grained image details during decoding, while the DM operates on a compressed latent representation that primarily captures high-level semantic structure The diffusion model uses a denoiser network trained to predict the denoised image from a noisy image at step t The reverse process for image generation is approximated by xt-1 ≈ xt − (σt − σt−1) · xt − Dθ(xt, σt) How the training signal is built The training data for the LDM is obtained from observations made with the Low Frequency Array (LOFAR) radio telescope Individual cutouts are extracted from publically available mosaic images issued with the third data release of the LOFAR Two-Metre Sky Survey (LoTSS DR3) Training images are obtained by dividing individual LoTSS mosaics into square-shaped tiles and saving them as arrays The LDM is trained on images of 512 px (12.8′), and the larger-sized cutouts facilitate random augmentations The inpainting mask is randomly varied during training with uniform probabilities between four possible shapes corresponding to sampling the entire image, sampling the right half, sampling the bottom half, and sampling the bottom right corner Where the gains come from The SWIIT-sampling procedure results in 152,098 cutouts of 1024 px side length The use of overlap sampling and blended decoding prevents edge-like decoding artifacts from appearing on the final map The SWIIT-sampling procedure demonstrates that no peaks are found at the exact sampling stride or multiples of it, showing no sign of periodic boundary artifacts on the sampled images The method is particularly suited for applications in which realistic survey-like images are desired, but less suitable for studies that require a clean separation between astrophysical sky emission and instrumental effects The model is capable of realistically modeling artifacts typically encountered around bright sources, as seen in Figure D.1 Data Availability The code implemented and used for this work is publicly available under https://github.com/tmartinezML/glori The authors thank the referee for a constructive report The authors acknowledge support by the Deutsche Forschungsgemeinschaft under Germany’s Excellence Strategy – EXC 2121 Quantum Universe – 390833306 and via the KISS consortium (05D23GU4) funded by the German Federal Ministry of Education and Research BMBF in the ErUM-Data action plan The references listed are Ascher, U.

Vera: M.

Jocelyn: & Petzold, L.

Vera: R.

Title and authors: Vera: So we're looking at this paper, "Generating radio continuum survey maps of arbitrary size with latent diffusion models." It’s about using these latent diffusion models to make realistic cutouts of radio survey maps that can be any size you want.

Jocelyn: And the authors are Vicánek Martínez and M. Brüggen, who are working on this at the Hamburger Sternwarte in Hamburg, Germany. It sounds like they're tackling the issue of creating synthetic data for observational astronomy when you need images of different scales.

Subrahmanyan: That’s a big deal because radio surveys generate an exponential amount of observational data, and having realistic simulations is crucial for testing new methods in data reduction and analysis, as mentioned on page one of this paper <ref:2609.06549#pg1>.

Vera: Exactly. The core idea is that diffusion models are great at making realistic images, but they usually have size limits or require a lot of computation. This work says they built a latent diffusion model to create these realistic radio survey map cutouts, and then developed a way to sample those maps no matter the size you need.

Jocelyn: And that sampling technique is called Sliding Window Iterative sampling, or SWIIT, which lets them break down the process into several steps for generating these huge sky surveys.

Subrahmanyan: What I find interesting is how they address that size limitation by using this iterative approach with context vectors and image inpainting to keep things consistent as they sample.

Vera: Right, so it’s not just about making one big map; it’s about being able to generate maps of arbitrary size while keeping the source locations and properties under control.

Jocelyn: And they use a catalog context vector that encodes information about where the sources are, their brightness, and their size. That gives them precise control over those properties when they generate the images.

Subrahmanyan: That control over source location and brightness is important because it allows for engineering that context vector to encode position, flux, and size information into the denoiser network as conditioning input.

Vera: And they even include information about sources in the immediate vicinity outside of the field of view to account for spillover artifacts, which helps make the iterative sampling coherent.

Title and authors: Jocelyn: That sounds like a clever way to handle those tricky edge cases when you're sliding that window around, ensuring you get consistent results.

Subrahmanyan: And from a cosmological side, it’s relevant because they are training this model on observations from the Low Frequency Array telescope, specifically the third data release of the LoTSS TwoMetre Sky Survey data.

Vera: That means they are grounding this generative model in real radio astronomy data from LOFAR to ensure it's modeling what we actually see. They used one hundred fifty-two thousand cutouts of one thousand twenty-four pixel side length for training, and they scaled the pixel values using a specific function gamma to handle the highly skewed distribution of radio map values.

Jocelyn: That scaling function is pretty important because it prepares the input images so that the large dynamic range of radio maps can be handled consistently across their entire dataset.

Subrahmanyan: So, for someone who only listens to the show, this means they’ve created a powerful tool that can generate realistic survey-like images at any size, which is something previous deep learning models couldn't do without these specific techniques.

Vera: It does mean they can generate realistic cutouts of five hundred twelve pixel side length with very precise control over where the sources are and how bright they are, which is a new capability for astronomical image generation compared to what we had before.

Jocelyn: And the results show that these LDM-sampled images replicate highly realistic source morphologies and associated imaging artifacts, which looks really good when you compare them to actual PyBDSF-detected sources.

Subrahmanyan: Replicating those artifacts is important because it means the synthetic data isn't just aesthetically pleasing; it has the necessary physical characteristics that matter for real science.

Vera: And they also developed this SWIIT sampling procedure which allows them to sample maps of arbitrary size, showing realistic results with only occasional artifacts that are seamless when you combine the individual samples.

Jocelyn: That seamless combination is key; it means the final map looks continuous even though it was built from many separate pieces sampled one by one.

Subrahmanyan: The method is particularly suited for applications where realistic survey-like images are desired, but they also point out a limitation: since the model was trained on observed radio survey maps, there isn't an explicit ground truth sky model against which they can compare the generated images directly.

Title and authors: Vera: That’s a fair caveat. And they also mentioned that the repeating-pattern artifacts observed in some sequentially generated examples might be a limitation of their sampling strategy.

Jocelyn: And another point is that this method is specific to LOFAR data, so applying it to surveys from other telescopes would require training on corresponding data from those instruments.

Subrahmanyan: They suggest that including minor axis and positional angle for each source could provide a more closely related input signal for better control over source size, which I think would make the model even more useful across different survey types.

Vera: So, to wrap up on this paper, they’ve created a latent diffusion model to generate realistic radio survey map cutouts of five hundred twelve pixels with precise control over source locations and brightness.

Jocelyn: And they developed SWIIT sampling to sample maps of arbitrary size, achieving realistic results with seamless combinations limited only by occasional artifacts.

Subrahmanyan: The implications are that this offers a way to generate synthetic data for testing new methods or even acting as an emulator for computationally expensive modeling steps, like data reduction and deconvolution.

Vera: It does mean they can generate novel realistic images that replicate source morphologies and associated imaging artifacts, which is valuable for generative data augmentation to improve classification models.

Jocelyn: We also see potential here for using this as an emulator for computationally expensive modeling steps, replacing the step of data reduction with training an LDM conditioned on a synthetic sky model.

Subrahmanyan: It’s useful because it provides a method that can be employed as an emulator for otherwise more computationally expensive modeling steps, given the availability of enough training images.

Vera: So we’ve covered the generation capability and the sampling technique in this paper by Vicánek Martínez and M. Brüggen's work on "Generating radio continuum survey maps of arbitrary size with latent diffusion models."

Jocelyn: That was a look at how they use these tools to tackle the size limitations of generative models and create consistent, large-scale synthetic radio map data.

Subrahmanyan: It’s a solid step in providing controlled environments for testing new methods in data analysis, which is where this kind of realistic simulation really fits into the larger picture of structure formation studies.

The paper's summary: Vera: So, this paper is about taking those latent diffusion models we use for generating images and figuring out how to make them work for radio survey maps that can be any size you want to sample.

Jocelyn: Exactly, they're using a technique called Sliding Window Iterative sampling, or SWIIT, to break down the process of making these huge sky surveys into smaller pieces we can actually manage.

Vera: It means they’ve managed to create realistic cutouts of five hundred twelve pixels with precise control over where the sources are and how bright they are, which is a new capability for generating astronomical images.

Jocelyn: And this SWIIT method lets them sample maps of any size, showing realistic results with only occasional artifacts that blend in when you put the pieces together.

Vera: The authors say they used a catalog context vector to encode the position, flux, and size information of each source to keep control over those properties during generation.

Jocelyn: That context vector is also used as conditioning input for the denoiser network, which helps train it to produce those specific features consistently.

Vera: They even include information about sources near the edge of the field of view in their sampling method, which they say accounts for spillover artifacts better during that sliding window process.

Jocelyn: That sounds like a really smart way to handle those tricky edge cases when you're moving the window around on a large map.

Vera: The big implication here is that this could be used to generate synthetic data for training other models, like those for classifying radio sources, by creating realistic examples of different source morphologies and artifacts.

Jocelyn: And they also suggest this method can act as an emulator for more computationally expensive steps in the pipeline, like doing all the heavy data reduction and deconvolution.

Vera: That’s a huge deal because it lets researchers test new modeling ideas without having to run those super long, expensive simulations every single time.

Jocelyn: It also shows that even with training on observed data from the LoTSS survey, you can build a model that generates synthetic images at arbitrary scales while maintaining good realism.

Vera: However, they do admit a limitation: since the model is trained on real radio maps, they don't have an explicit ground truth sky model to compare against for validation.

Jocelyn: And they also point out that the repeating-pattern artifacts seen in some of their sequential samples might be a sign of a weakness in the sampling strategy itself.

Vera: Plus, this specific method is tuned for LOFAR data, so applying it to observations from different telescopes would require retraining it on new data from those instruments.

Jocelyn: They suggest that adding more details about the source's minor axis and positional angle could give them a better input signal for even tighter control over the source size.

Vera: So, we’ve got this way to create high-fidelity, controllable synthetic radio map cutouts of any size using latent diffusion models and an iterative sampling approach.

Jocelyn: It’s a really powerful tool for generative augmentation and as a way to model complex observational steps efficiently.

Vera: And it opens the door for testing new analysis techniques in environments where you can perfectly control the input data characteristics.

Jocelyn: Next, we're going to look at some other papers on how we handle instrumental polarization in CMB experiments.

The paper's improvements: Tom: So, we're talking about what they suggest to make this map generation even better than it is right now.

Vera: They mention that since this model was trained on real radio survey maps, there isn't a perfect ground truth sky model for direct comparison.

Jocelyn: That’s true, and they note that the repeating pattern artifacts we saw in some of their samples might just be a limitation of how they are sliding the window.

Vera: To fix that, they suggest including more source details into the input signal, specifically adding information about each source's minor axis and its positional angle.

Jocelyn: That makes sense because knowing those things gives them a much better handle on controlling the size of each source when they generate the map.

Vera: If you know that, you can engineer a catalog context vector that encodes position, brightness, and size with much more precision than just using basic position and flux.

Jocelyn: And because they are including information about sources right outside the field of view too, it should help them model those spillover artifacts during the sliding window sampling process more correctly.

Vera: That suggests a path forward for improving control over source properties when generating these synthetic images.

Jocelyn: It means that if you want to generate a map where you can precisely dictate the size of every feature, you need to feed the model richer information about those features.

Vera: So, we’re moving from just sampling arbitrary sizes to having much tighter control over the physical characteristics of what gets sampled.

Jocelyn: The authors also hint that they are looking at consistency models as an alternative way to reduce the high computational cost associated with running all those sequential denoising steps.

Vera: That’s a big thing because running the model step by step for every single part of the map is what makes it slow, and consistency models promise a faster way to generate coherent results.

Jocelyn: If they get that right, it could make these kinds of large-scale simulations much more practical for real-time analysis or quicker testing.

Vera: It shifts the focus from just generating an image to finding a more efficient way to model the entire generation process itself.

Jocelyn: So, while the current SWIIT sampling is a solid way to get arbitrary sizes now, their suggestions point toward a future where control and speed are both improved.

Vera: It’s about making sure that when we generate these synthetic maps, we aren't just getting something visually realistic but something physically controllable.

Jocelyn: Next up, we're going to talk about how instrumental polarization in CMB experiments is being studied right now.

Conclusion: Tom: So we're wrapping up on "Generating radio continuum survey maps of arbitrary size with latent diffusion models." It sounds like they’ve shown us how to sample these massive sky surveys in a way that’s truly flexible.

Vera: Right, and the main thing is getting those realistic cutouts of five hundred twelve pixels with real control over where the sources are located and how bright they are.

Jocelyn: And they showed that their sliding window sampling technique, SWIIT, lets them build those huge maps piece by piece without major artifacts when you combine them.

Subrahmanyan: From a cosmological standpoint, being able to generate these synthetic data sets at arbitrary sizes is useful because it lets us test how different source distributions affect our simulations of structure formation.

Vera: And the control over source properties via that catalog context vector is key for making sure the synthetic maps actually represent what we might see in real sky surveys.

Jocelyn: It’s a powerful way to create training data for other machine learning models, which is important since we need realistic inputs to teach those algorithms how to find faint signals.

Subrahmanyan: If these models can generate images that look physically plausible, it means we can use them as an emulator for more expensive modeling steps in the universe, like simulating the full data reduction process.

Vera: Exactly, so this method serves as a way to test new data processing techniques without having to run those massive simulations every time.

Jocelyn: And even though they mentioned that it’s specific to LOFAR data and there are some repeating pattern artifacts, the core idea of flexible sampling is definitely something we need to keep exploring.

Subrahmanyan: That limitation on applying it elsewhere is a practical one; you need the right data for the right instrument, but the method itself points toward a direction for more versatile generative modeling in astronomy.

Vera: So, to recap, this paper on "Generating radio continuum survey maps of arbitrary size with latent diffusion models" gives us a robust tool for creating controllable synthetic radio map cutouts and sampling them at any scale.

Jocelyn: It’s a strong step toward making generative models useful for observational astronomy by giving us realistic data that we can manipulate.

Subrahmanyan: It’s an important technical achievement because it allows us to test the robustness of new analysis methods in environments where we can perfectly control the input data characteristics.

Vera: We've seen how they handle size and control, and now we know there's a path for more efficient, controllable synthetic image generation.

Jocelyn: Next time, we’re looking at how instrumental polarization in CMB experiments is being studied to make sure we aren't confusing real signals with noise.

T. Vicánek Martíneze, M. Br뾎ggen

Hamburger Sternwarte, University of Hamburg

astro-ph.IM

Submitted: 2026-09-06

Updated: 2026-09-06

Journal ref: A&A 715, A248 (2026)

DOI: 10.1051/0004-6361/202660213

Code: https://github.com/tmartinezML/glori

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 89/100

The gist: Generating radio continuum survey maps of arbitrary size with latent diffusion models implements a latent diffusion model (LDM) to synthesize realistic radio survey map cutouts and a technique called…

Key concepts

Latent Diffusion Model (LDM)
An LDM combines a Vector Quantized Variational Autoencoder (VQ-VAE) with a diffusion model. The VQ-VAE compresses images into a smaller latent space, and the diffusion model then operates on this compressed representation to generate realistic map cutouts.
Sliding Window Iterative Sampling (SWIIT)
SWIIT is a sampling technique that breaks down large map generation into steps. It samples small patches within a sliding window, moving the window by half the image size each time. Overlapping parts are used to ensure consistency when stitching these individual samples together to form an arbitrarily large map.
Context Vector
A context vector is a set of numerical inputs that guides the generation process. In this work, it encodes crucial information about radio sources, including their position, brightness (flux), and size. This allows for precise control over where and how bright the generated sources appear in the final map.
Pixel Value Scaling ($\gamma$ function)
Radio survey data has a very wide range of pixel values. A scaling function is used to transform these original, highly skewed values into a distribution that is close to zero-mean and unit variance. This preprocessing step makes the data suitable for training the generative model consistently.

Terminology

Summary

Generating radio continuum survey maps of arbitrary size with latent diffusion models implements a latent diffusion model (LDM) to synthesize realistic radio survey map cutouts and a technique called Sliding Window Iterative (SWIIT) sampling to sample these maps of arbitrary size The method addresses the limitations of fixed-size generative models by enabling the generation of large-scale synthetic radio maps with control over source locations and properties

Model Architecture and Training

The LDM is composed of a vector-quantized variational autoencoder (VQ-VAE) and a diffusion model operating on the resulting latent representations The VQ-VAE uses an encoder to project images into a smaller latent space, which is then processed by the diffusion model The VQ-VAE architecture features three resolution levels and a latent dimension of 3 channels Training the VQ-VAE involves several loss terms, including reconstruction loss, discriminator loss, and commitment loss The diffusion model denoiser network is implemented as a U-Net that incorporates conditional inputs such as the noise level and the inpainting context The training schedule involves several stages, including an initial run for 2 × 10 5 iterations and a second stage with gradient accumulation for another 2 × 10 5 iterations

Sampling Technique

The SWIIT-sampling technique is developed to break down the process of sampling large sky survey maps into several individual steps This approach works by sampling individual patches within a sliding window moved with a stride of half the image size, passing overlapping parts as additional input for consistency The procedure involves sequentially sampling the entire image, then moving the window to sample only the right half, and so on until the desired map width is reached The final latent map is decoded into image space using a sliding-window convolutional fashion with a stride of half the latent image size

Performance and Control

The results show that realistic cutouts of 512 pixel side length can be generated with precise control over source locations and brightness Sampling of arbitrary-sized maps is made possible, showing realistic results with seamless combination of individual samples limited only by occasional sampling artifacts The LDM-sampled images replicate highly realistic source morphologies and associated imaging artifacts Control over source location and brightness properties is facilitated by engineering a catalog context vector that encodes position, flux, and size information

Limitations and Future Directions

A limitation of the present work is that since the model is trained on observed radio survey maps, there is no explicit ground-truth sky model against which the generated images can be compared The repeating-pattern artifacts observed in some sequentially generated examples may indicate a limitation of the proposed sampling strategy Furthermore, the method is specific to LOFAR data, and application to surveys from other telescopes would require training on corresponding data from those instruments The authors suggest that including minor axis and positional angle for each source could provide a more closely related input signal for better control over source size

Data Preparation

The training data is obtained from observations made with the Low Frequency Array (LOFAR) radio telescope, specifically from the third data release of the LOFAR Two-Metre Sky Survey (LoTSS DR3) Individual cutouts are extracted from these mosaics, and pixel values are scaled using a function γ to transform the asymmetric distribution into one close to zero-mean and unit variance Information on source positions, fluxes, and sizes is encoded into a catalog context vector which is passed as conditioning input to the denoiser network The training set consists of 152,098 cutouts of 1024 px side length

Summary and Conclusions

In this work, we implemented a generative model for synthesizing realistic individual radio survey map images of 512 px side length, a size that is to our knowledge unprecedented for generative models in the context of astronomical image generation Further, we developed SWIIT-sampling, a novel technique for sampling radio sky survey maps of arbitrary size through iterative use of an LDM trained for image inpainting By designing a custom function for scaling the pixel values of training images, the large dynamic range of radio maps can be dealt with in a consistent way over the entire dataset This makes radio observational data suitable for computer vision tasks while still conserving universal flux information By engineering a context vector that encodes spatial information of radio sources present on the images, as well as their brightness and size properties, we facilitate effective control over sources present on the sampled images By also including information of sources in the immediate vicinity outside of the FOV, we account for a correct modeling of spillover artifacts and thereby allow for coherent iterative sampling in a sliding-window approach This makes it possible to sample realistic images at arbitrary sizes, circumventing current size limitations of deep learning models for image generation The code implemented and used for this work is publicly available under https://github.com/tmartinezML/glori The authors acknowledge support by the Deutsche Forschungsgemeinschaft under Germany’s Excellence Strategy – EXC 2121 Quantum Universe – 390833306 and via the KISS consortium (05D23GU4) funded by the German Federal Ministry of Education and Research BMBF in the ErUM-Data action plan The paper concludes that while results show remarkable realism and quality, the proposed method works at the expense of high computational cost at both training and inference time The authors also note that consistency models are a promising alternative to alleviate the computational cost associated with sequential denoising procedures The paper is applicable to generating any kinds of large-scale mock image data, given the availability of enough training images The authors also note that the quality of the source catalogs used for training is a limiting factor in this regard The paper provides a method that can be employed as an emulator for otherwise more computationally expensive modeling steps

Pruning redundant reasoning tokens

The authors describe the LDM as combining DMs with VAEs, where the DM operates inside the latent space of a VQ-VAE The VQ-VAE is responsible for reconstructing fine-grained image details during decoding, while the DM operates on a compressed latent representation that primarily captures high-level semantic structure The diffusion model uses a denoiser network trained to predict the denoised image from a noisy image at step t The reverse process for image generation is approximated by xt-1 ≈ xt − (σt − σt−1) · xt − Dθ(xt, σt)

How the training signal is built

The training data for the LDM is obtained from observations made with the Low Frequency Array (LOFAR) radio telescope Individual cutouts are extracted from publically available mosaic images issued with the third data release of the LOFAR Two-Metre Sky Survey (LoTSS DR3) Training images are obtained by dividing individual LoTSS mosaics into square-shaped tiles and saving them as arrays The LDM is trained on images of 512 px (12.8′), and the larger-sized cutouts facilitate random augmentations The inpainting mask is randomly varied during training with uniform probabilities between four possible shapes corresponding to sampling the entire image, sampling the right half, sampling the bottom half, and sampling the bottom right corner

Where the gains come from

The SWIIT-sampling procedure results in 152,098 cutouts of 1024 px side length The use of overlap sampling and blended decoding prevents edge-like decoding artifacts from appearing on the final map The SWIIT-sampling procedure demonstrates that no peaks are found at the exact sampling stride or multiples of it, showing no sign of periodic boundary artifacts on the sampled images The method is particularly suited for applications in which realistic survey-like images are desired, but less suitable for studies that require a clean separation between astrophysical sky emission and instrumental effects The model is capable of realistically modeling artifacts typically encountered around bright sources, as seen in Figure D.1

Data Availability

The code implemented and used for this work is publicly available under https://github.com/tmartinezML/glori The authors thank the referee for a constructive report The authors acknowledge support by the Deutsche Forschungsgemeinschaft under Germany’s Excellence Strategy – EXC 2121 Quantum Universe – 390833306 and via the KISS consortium (05D23GU4) funded by the German Federal Ministry of Education and Research BMBF in the ErUM-Data action plan The references listed are Ascher, U. M. & Petzold, L. R. 1998, in Computer methods for ordinary differential equations and differential-algebraic equations Best, P. p. 0

Where the paper says it

Given the richness of such observations that cover wide angular scales at high resolution and sensitivities, these advances open substantial opportunities for discoveries in diverse areas, including the evolution of galaxies and active galactic nuclei (e. Best et al.

N., Ker, L. M., Simpson, C., Rigby, E. E., & Sabater, J. 2014, MNRAS Bonaldi, A. p. 0
Where the paper says it

Best et al. 2014; Wilman et al.

, An, T., Brüggen, M. p. 2
Where the paper says it

Vicánek Martínez and M. Brüggen : Generating radio continuum survey maps of arbitrary size with latent di ˇ ffusion models at step t can be approximated as xt−1 ≈ xt − (σt − σt−1) · xt − Dθ(xt , σt) σt .

, et al. 2021 Cassano, R. p. 0
Where the paper says it

2015), as well as the study of magnetic fields and the intergalactic medium (e. Cassano et al.

, Brunetti, G., Röttgering, H. J. A., & Brüggen, M. p. 2
Where the paper says it

Vicánek Martínez and M. Brüggen : Generating radio continuum survey maps of arbitrary size with latent di ˇ ffusion models at step t can be approximated as xt−1 ≈ xt − (σt − σt−1) · xt − Dθ(xt , σt) σt .

p. 2
Where the paper says it

Vicánek Martínez and M. Brüggen : Generating radio continuum survey maps of arbitrary size with latent di ˇ ffusion models at step t can be approximated as xt−1 ≈ xt − (σt − σt−1) · xt − Dθ(xt , σt) σt .

p. 2
Where the paper says it

Vicánek Martínez and M. Brüggen : Generating radio continuum survey maps of arbitrary size with latent di ˇ ffusion models at step t can be approximated as xt−1 ≈ xt − (σt − σt−1) · xt − Dθ(xt , σt) σt .

2010 Chen, H. p. 1
Where the paper says it

While successful, their application is limited by practical constraints (see e. Chen et al.

, Xiang, Q., Hu, J., et al. 2025 Cornwell, T. p. 0
Where the paper says it

2009; Cornwell et al. 2025), their intrinsic source morphology models are restricted to simple Gaussian shapes or a limited library of observed sources, and thus they do not fully capture the statistical diversity of extended radio emission.

J., Wortmann, P., Nikolic, B., Wang, F., & Stolyarov, V. 2025 Dewdney, P. p. 0
Where the paper says it

2009; Cornwell et al. 2025), their intrinsic source morphology models are restricted to simple Gaussian shapes or a limited library of observed sources, and thus they do not fully capture the statistical diversity of extended radio emission.

E., Hall, P. J., Schilizzi, R. T., & Lazio, T. J. L. W. 2009 Dhariwal, P. p. 0
Where the paper says it

2023; Dulwich et al. 2009; Cornwell et al.

, & Nichol, A. p. 0
Where the paper says it

Recently, deep learning–based generative approaches have emerged as powerful tools for synthesizing image data. In particular, diffusion models (DMs) have demonstrated remarkable performance in high-fidelity generation of natural images (Dhariwal & Nichol 2021; Karras et al.

2021 Drozdova, M. p. 1
Where the paper says it

2024), radio-astronomical image reconstruction (Wang et al. 2023; Drozdova et al.

, Kinakh, V., Bait, O., et al. 2024 Dulwich, F. p. 0
Where the paper says it

2021; Hartley et al. 2023; Dulwich et al.

, Mort, B. J., Salvini, S., Zarb Adami, K., & Jones, M. E. 2009 Hartley, P. p. 0
Where the paper says it

While existing approaches for simulating radio observations, such as the SKA Science Data Challenges, OSKAR, and RASCIL, offer detailed models of instrumental and telescope effects (Bonaldi et al. 2021; Hartley et al.

, Bonaldi, A. p. 0
Where the paper says it

Finally, radio sources are not uniformly distributed across the sky but exhibit spatial correlations driven by large-scale structure and influenced by survey selection effects, which must be faithfully reproduced to generate statistically representative observational data. While existing approaches for simulating radio observations, such as the SKA Science Data Challenges, OSKAR, and RASCIL, offer detailed models of instrumental and telescope effects (Bonaldi et al.

, Braun, R., et al. 2023 Ho, J. p. 0
Where the paper says it

Galaxies: general - Methods: data analysis - Techniques: image processing - Surveys - Radio continuum: Galaxies 1. Introduction Following the continuously improving capacities of radio astronomical instruments experienced over the past decades, radio sky surveys provide observational data at exponentially increasing volumes (Smith & Geach 2023).

& Salimans, T. p. 2
Where the paper says it

Guided diffusion The NN can be trained with additional inputs c of any shape, such as class labels or parameters that characterize the information contained on the image. This use of additional input information is referred to as conditioning, and DMs achieve high quality in conditional sampling through a technique called classifier-free guidance (Ho & Salimans 2022).

2022 Huh, M. p. 6
Where the paper says it

Vector quantization happens inbetween. For training the VQ-VAE, we adopt several strategies from Huh et al.

, Cheung, B., Agrawal, P., & Isola, P. 2023 Karras, T. p. 0
Where the paper says it

Recently, deep learning–based generative approaches have emerged as powerful tools for synthesizing image data. In particular, diffusion models (DMs) have demonstrated remarkable performance in high-fidelity generation of natural images (Dhariwal & Nichol 2021; Karras et al.

, Aittala, M., Aila, T., & Laine, S. 2022 Kingma, D. p. 1
Where the paper says it

The training loss, apart from enforcing accurate image reconstruction, typically constraints the representations of the training images to follow a distribution in latent space that can easily be sampled from. This is done in different ways, giving rise to multiple variants of the model class, like the widely used VAE (Kingma & Welling 2019) or the vector-quantized variational autoencoder (VQ-VAE, van den Oord et al.

P. & Welling, M. p. 1
Where the paper says it

The training loss, apart from enforcing accurate image reconstruction, typically constraints the representations of the training images to follow a distribution in latent space that can easily be sampled from. This is done in different ways, giving rise to multiple variants of the model class, like the widely used VAE (Kingma & Welling 2019) or the vector-quantized variational autoencoder (VQ-VAE, van den Oord et al.

2019 Ma, C. p. 1
Where the paper says it

The training loss, apart from enforcing accurate image reconstruction, typically constraints the representations of the training images to follow a distribution in latent space that can easily be sampled from. This is done in different ways, giving rise to multiple variants of the model class, like the widely used VAE (Kingma & Welling 2019) or the vector-quantized variational autoencoder (VQ-VAE, van den Oord et al.

, Sun, Z., Jing, T., et al. 2026 Mishra, S. p. 0
Where the paper says it

Astronomy & Astrophysics manuscript no. aa60213-26 ©ESO 2026 Wednesday 9th September, 2026 Generating radio continuum survey maps of arbitrary size with latent diffusion models T.

, Trotta, R., & Viel, M. 2026 Mohan, N. p. 3
Where the paper says it

The mosaics have a resolution of 6′′ above declination 10◦ and 9′′ resolution below. The data release also features a source catalog, obtained with the Python Blob Detector and Source Finder (PyBDSF, Mohan & Rafferty 2015), which we use to create our catalog context arrays.

& Rafferty, D. p. 3
Where the paper says it

The mosaics have a resolution of 6′′ above declination 10◦ and 9′′ resolution below. The data release also features a source catalog, obtained with the Python Blob Detector and Source Finder (PyBDSF, Mohan & Rafferty 2015), which we use to create our catalog context arrays.

2015 Nicolaas, R. p. 1
Where the paper says it

2024), and for removal of noise and artifacts (Waldmann et al. 2023; Nicolaas et al.

, Caron, S., Stoppa, F., et al. 2026 Perez, E. p. 7
Where the paper says it

Alongside the input image, the U-Net receives conditional inputs of different kinds. The diffusion noise level σ is passed as a scalar value to a fully-connected network (FCN) before being injected into each of the CNN blocks between the first and second CNN layer through feature-wise linear modulation (FiLM, Perez et al.

, Strub, F., De Vries, H., Dumoulin, V., & Courville, A. 2018 Potevineau, R. p. 1
Where the paper says it

2024; Ma et al. 2026; Potevineau ˇ et al.

, Tolley, E., & Etsebeth, V. 2026 Qiu, Z. p. 0
Where the paper says it

Astronomy & Astrophysics manuscript no. aa60213-26 ©ESO 2026 Wednesday 9th September, 2026 Generating radio continuum survey maps of arbitrary size with latent diffusion models T.

, Wang, Z., Zheng, B., et al. 2025 Reddy, P. p. 1
Where the paper says it

2026; Scognamiglio et al. 2025), image deconvolution and super-resolution (Reddy et al.

, Toomey, M. W., Parul, H., & Gleyzer, S. 2024 Rombach, R. p. 1
Where the paper says it

To this end, we employ a variant of DMs referred to as latent diffusion models (LDMs). First introduced in Rombach et al.

, Blattmann, A., Lorenz, D., Esser, P., & Ommer, B. 2021 Ronneberger, O. p. 5
Where the paper says it

4: Illustration of the cutout extraction for a single LoTSSDR3 mosaic, showing the tiling pattern (left) and the resulting cutouts after pixel scaling (right). Model architecture and training The encoder-decoder architecture used for the VQ-VAE is nearly identical to the U-Net (Ronneberger et al.

, Fischer, P., & Brox, T. 2015 Schwarz, D. p. 0
Where the paper says it

Schwarz et al. 2015), as well as the study of magnetic fields and the intergalactic medium (e.

J., Bacon, D. J., Chen, S., et al. 2015 Shimwell, T. p. 0
Where the paper says it

Schwarz et al. 2015), as well as the study of magnetic fields and the intergalactic medium (e.

W., Hale, C. L., Best, P. N., et al. 2025 Shimwell, T. p. 8
Where the paper says it

To quantify the similarity in source properties, we run PyBDSF, using the same settings employed for the generation of the DR3 source catalog, see Shimwell et al. (2025) for a detailed description of the procedure.

W., Hardcastle, M. J., Tasse, C., et al. 2026 Smith, M. p. 0
Where the paper says it

Galaxies: general - Methods: data analysis - Techniques: image processing - Surveys - Radio continuum: Galaxies 1. Introduction Following the continuously improving capacities of radio astronomical instruments experienced over the past decades, radio sky surveys provide observational data at exponentially increasing volumes (Smith & Geach 2023).

J. & Geach, J. p. 0
Where the paper says it

Galaxies: general - Methods: data analysis - Techniques: image processing - Surveys - Radio continuum: Galaxies 1. Introduction Following the continuously improving capacities of radio astronomical instruments experienced over the past decades, radio sky surveys provide observational data at exponentially increasing volumes (Smith & Geach 2023).

E. 2023 Smith, M. p. 0
Where the paper says it

Galaxies: general - Methods: data analysis - Techniques: image processing - Surveys - Radio continuum: Galaxies 1. Introduction Following the continuously improving capacities of radio astronomical instruments experienced over the past decades, radio sky surveys provide observational data at exponentially increasing volumes (Smith & Geach 2023).

J., Geach, J. p. 0
Where the paper says it

Galaxies: general - Methods: data analysis - Techniques: image processing - Surveys - Radio continuum: Galaxies 1. Introduction Following the continuously improving capacities of radio astronomical instruments experienced over the past decades, radio sky surveys provide observational data at exponentially increasing volumes (Smith & Geach 2023).

E., Jackson, R. A., et al. 2022 Song, Y. p. 0
Where the paper says it

In particular, diffusion models (DMs) have demonstrated remarkable performance in high-fidelity generation of natural images (Dhariwal & Nichol 2021; Karras et al. 2022, 2024).

, Dhariwal, P. p. 0
Where the paper says it

Recently, deep learning–based generative approaches have emerged as powerful tools for synthesizing image data. In particular, diffusion models (DMs) have demonstrated remarkable performance in high-fidelity generation of natural images (Dhariwal & Nichol 2021; Karras et al.

, Chen, M., & Sutskever, I. 2023 Sortino, R. p. 1
Where the paper says it

This allows for a reduction of the DMs computational demand by effectively compressing images to a smaller size through use of the VAE, while still preserving relevant details and thereby assuring high sample quality for large image sizes. In the context of radio astronomy, the only use of LDMs is by Sortino et al.

, Cecconello, T., De Marco, A., et al. 2024 Spagnoletti, A. p. 1
Where the paper says it

2024; Shan et al. 2025; Spagnoletti et al.

, Boucaud, A., Huertas-Company, M., Kabalan, W., & Biswas, B. 2024 van den Oord, A. p. 1
Where the paper says it

The training loss, apart from enforcing accurate image reconstruction, typically constraints the representations of the training images to follow a distribution in latent space that can easily be sampled from. This is done in different ways, giving rise to multiple variants of the model class, like the widely used VAE (Kingma & Welling 2019) or the vector-quantized variational autoencoder (VQ-VAE, van den Oord et al.

, Vinyals, O., & Kavukcuoglu, K. 2017 van Haarlem, M. p. 1
Where the paper says it

This is done in different ways, giving rise to multiple variants of the model class, like the widely used VAE (Kingma & Welling 2019) or the vector-quantized variational autoencoder (VQ-VAE, van den Oord et al. (2017)), which we use in this work.

P., Wise, M. W., Gunst, A. W., et al. 2013 Vicánek Martínez, T. p. 0
Where the paper says it

aa60213-26 ©ESO 2026 Wednesday 9th September, 2026 Generating radio continuum survey maps of arbitrary size with latent diffusion models T. Vicánek Martínez ˇ 1 and M.

, Edler, H. W., & Brüggen, M. 2025 Vicánek Martínez, T. p. 0
Where the paper says it

aa60213-26 ©ESO 2026 Wednesday 9th September, 2026 Generating radio continuum survey maps of arbitrary size with latent diffusion models T. Vicánek Martínez ˇ 1 and M.

, Baron Perez, N. p. 7
Where the paper says it

Alongside the input image, the U-Net receives conditional inputs of different kinds. The diffusion noise level σ is passed as a scalar value to a fully-connected network (FCN) before being injected into each of the CNN blocks between the first and second CNN layer through feature-wise linear modulation (FiLM, Perez et al.

, & Brüggen, M. 2024 Waldmann, I. p. 1
Where the paper says it

2023; Drozdova et al. 2024), and for removal of noise and artifacts (Waldmann et al.

, Rocchetto, M., & Debczynski, M. 2023 Wang, R. p. 1
Where the paper says it

2024), radio-astronomical image reconstruction (Wang et al. 2023; Drozdova et al.

, Chen, Z., Luo, Q., & Wang, F. p. 1
Where the paper says it

2025; Spagnoletti et al. 2024), radio-astronomical image reconstruction (Wang et al.

2023 Wilman, R. p. 0
Where the paper says it

Best et al. 2014; Wilman et al.

J., Miller, L., Jarvis, M. J., et al. 2008

Appendix A: Pixel value scaling

The pixel values of the LoTSS mosaics follow a highly skewed and asymmetric distribution that spans several orders of magnitude To prepare the input images for efficient model training, we apply a scaling designed to make the pixel value distribution approximately have zero-mean and unit variance The function γ employed to obtain the scaled pixel values from the original values x is defined as γ±(x) = a± s· ln (a± · x + 1)! The inverse scaling function is given by γ−1±(y) = a± s· e y/a± − 1

Appendix B: Optimal tiling pattern for cutout extraction

The optimal pattern for dividing any LoTSS-DR3 mosaic into square-shaped cutouts of a side length L is determined by calculating the maximum valid mosaic width W and then dividing this patch horizontally into h bins For each column, the maximum valid height H is calculated between the lower and upper bound of valid pixels The procedure then divides the column vertically into v squares of side length L

Appendix C: Model architecture and training details

The VQ-VAE architecture is illustrated in Figure 7, featuring three resolution levels and a latent dimension of 3 channels The VQ-VAE is optimized with a linear combination of loss terms LVQ-VAE = αRecLRec + αRec,unscLRec,unsc + αGenLGen + αCommitLCommit The DM denoiser U-Net architecture features residual connections between the encoder and decoder parts at levels of equal resolution The denoiser Dθ is trained with L2 reconstruction loss, defined as L = cout(σ) − 2 · Dθ(x + z, σ) − x squared / 2

**Appendix C.

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

  1. Latent diffusion models can generate realistic cutouts of 512 pixel side length with precise control over source locations and brightness, which is a new capability for astronomical image generation compared to previous limitations.

  2. A technique called Sliding Window Iterative sampling (SWIIT) allows for the sampling of radio survey maps of arbitrary size by using a sliding window with a stride of half the image size, facilitating seamless combination of individual samples and limited only by occasional sampling artifacts.

  3. The system can generate novel realistic images that replicate source morphologies and associated imaging artifacts, as evidenced by the qualitative visual similarity to original cutouts in comparison to PyBDSF-detected sources.

  4. The model can be used for generative data augmentation to improve un- or self-supervised learning of source classification models, allowing algorithms to test edge cases like very extended sources or highly clustered fields.

  5. The method can serve as an emulator for computationally expensive modeling steps, such as replacing the highly computational step of data reduction and deconvolution with training an LDM conditioned on a synthetic sky model.

Abstract

Radio surveys provide exponentially increasing amounts of observational data, which requires the development of novel data reduction and analysis methods. To this end, realistic simulations of observational data provide controlled environments for testing and developing new methods. Diffusion models excel at synthesizing realistic image data, but are limited by size constraints and high computational demands. We implemented a latent diffusion model (LDM) to synthesize realistic radio survey map cutouts. Further, we developed a technique for sampling radio survey maps of arbitrary size with the LDM. This is done by sampling in multiple sequential steps, whereby consistency is provided through context vectors and image inpainting. We trained our model on observations from the LOFAR telescope obtained from the third data release of the LOFAR Two-Metre Sky Survey. We implemented a LDM composed of a vector-quantized variational autoencoder and a diffusion model operating on the resulting latent representations. We added information on source locations, brightness and sizes by extracting parameters from the public source catalog and encoding them into a vector format passed as context to the LDM. We trained the model for image inpainting to facilitate continuous sequential sampling. We were able to generate realistic cutouts of 512 pixel side length, corresponding to 5.7 arcmin, with precise control over source locations and brightness. Control over source extension was achieved at lower precision. Sampling of arbitrary-sized maps was made possible, showing realistic results with seamless combination of individual samples and limited only by occasional sampling artifacts.

Sources

Related papers