From Spaceborne to Airborne: SAR Image Synthesis Using Foundation Models for Multi-Scale Adaptation

Elise Colin; Nathan Letheule; Nicolas Trouve; Olivier Leveque; Solene Debuysere

REVIEW 4 major objections 5 minor 1 cited by

From Spaceborne to Airborne: SAR Image Synthesis Using Foundation Models for Multi-Scale Adaptation

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper claims that a spatially conditioned latent diffusion foundation model can turn satellite SAR images into 40 cm airborne-like imagery and make physics-based simulated SAR images more realistic.

desk verdict A reasonable engineering pipeline for SAR domain adaptation, but the central effectiveness claim rests on qualitative figures alone and needs quantitative backing before it should be trusted. read the letter →

arxiv 2505.03844 v2 pith:XZ6NFV3F submitted 2025-05-05 eess.IV cs.AIcs.CV

Solene Debuysere , Nicolas Trouve , Nathan Letheule , Olivier Leveque , Elise Colin This is my paper · ORCID

classification eess.IVcs.AIcs.CV

keywords syntheticapertureradarSARimagesynthesislatentdiffusionmodelsStableXLControlNetmulti-scaleupscalingdataaugmentationremotesensing

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper aims to show that a large pre-trained text-to-image diffusion model can be adapted to Synthetic Aperture Radar (SAR) through LoRA fine-tuning and spatial conditioning, converting relatively low-resolution satellite SAR images into higher-resolution, airborne-style SAR images. If true, this would give the remote-sensing community a data augmentation path that bypasses costly airborne acquisition campaigns. The authors also apply the same pipeline to simulated SAR images from a physics-based radar simulator, arguing it makes simulated imagery more realistic. The work stands on a new dataset of roughly 110,000 40 cm-resolution SAR image samples built from 15 years of airborne X-band acquisitions.

What carries the argument

The central mechanism is the 'creative upscaling pipeline': a progressive latent-space upscaling chain inside Stable Diffusion XL (SDXL), a 3.5-billion-parameter latent diffusion model, in which two ControlNet modules apply Canny edge and Tile texture filters as spatial conditioning at successive resolution stages. The SDXL backbone is adapted to SAR through Low-Rank Adaptation (LoRA) fine-tuning at three resolutions (160 cm, 80 cm, and 40 cm), producing resolution-specific models that each refine one stage of the pipeline. The spatial condition term extracted from the input image enters the reverse diffusion mean, which is what lets the pipeline preserve global structure while the generative model adds new high-resolution detail.

What would settle it

Take a satellite SAR scene for which a real 40 cm airborne SAR image of the same ground area exists, run the full pipeline on the satellite image, and compare the output against the real airborne image on three quantitative axes: speckle statistics (for instance, the distribution of backscatter intensity), edge and structure preservation (via a structural similarity index or a SAR-specific edge metric), and radiometric calibration. If the synthesized image is statistically indistinguishable from the real airborne acquisition on these axes, the claim survives; if the output shows optical-image-like texture or fails the SAR statistics test, the central claim would be refuted.

Watch

Extended reading notes

Core claim

The paper argues that a multi-stage latent upscaling pipeline built on Stable Diffusion XL and ControlNet modules can transform 80 cm-resolution satellite SAR images into 40 cm-resolution images with an airborne SAR appearance, and can inject realistic fine texture into simulated images from a physics-based radar simulator. The transformation is achieved by fine-tuning the 3.5-billion-parameter latent diffusion model with LoRA on a curated dataset of roughly 110,000 40 cm-resolution SAR samples drawn from 15 years of airborne X-band acquisitions, then guiding each of two successive upscaling steps (512-by-512 to 1024-by-1024, then to 2048-by-2048) with Canny and Tile ControlNet modules that act as spatial conditions during denoising. The authors present this as the first application of such a foundation-model approach to satellite-to-airborne SAR adaptation.

Load-bearing premise

The load-bearing premise is that ControlNet modules trained on optical images can still provide useful structural guidance for SAR images during denoising, even though SAR's speckle and backscatter statistics are unlike anything those modules have seen; the paper itself concedes that this mismatch likely degraded performance.

Editorial extensions

If this is right

If the claim holds, satellite SAR imagery becomes a scalable source of training data for 40 cm-resolution applications, since the pipeline converts cheap and widely available spaceborne images into airborne-style samples without new acquisitions.
The same pipeline gives physics-based SAR simulation a realism bridge, meaning simulated scenes could be used to train or test algorithms that currently require real high-resolution data.
The approach extends the authors' earlier 512-to-2048 progressive upscaling result to a foundation-model setting, showing that pre-trained optical diffusion models can be repurposed for radar through LoRA plus spatial conditioning.
The curated 110,000-sample, 40 cm-resolution dataset becomes a resource for further fine-tuning, whether or not the authors release it.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural next experiment, which the paper does not run, is to fine-tune ControlNet on SAR images rather than using optical-tuned modules; if the authors' own diagnosis is right, SAR-specific ControlNet should improve structure preservation and reduce the large structural modifications they observed at higher control strengths.
Because the pipeline is resolution-agnostic after LoRA fine-tuning, the same architecture could be tested for other sensor gaps, such as L-band or C-band satellite-to-airborne transfer, or for optical-to-SAR cross-modal synthesis.
Quantitative evaluation is the missing piece the paper leaves implicit: a detection or segmentation benchmark trained on the synthesized 40 cm images and tested on real airborne data would show whether the generated textures are genuinely useful for downstream remote sensing or only visually plausible.
The pairing of SAR images with optical captions suggests the text prompt itself carries scene semantics; ablating the caption would test how much of the output fidelity comes from the text condition versus the spatial ControlNet condition.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You asked about 2505.03844. My quick take: it's a plausible engineering extension of the same group's earlier diffusion-based SAR synthesis work, now using SDXL plus ControlNet and LoRA. That combination is new, and it's built on a serious dataset effort: 110K images from 15 years of SETHI airborne acquisitions. Credit where due: the authors are transparent about a major limitation in their own discussion, which is more than many papers do.

What's actually good: the multi-stage pipeline (latent upscale, then ControlNet-guided refinement at 80cm and 40cm) is a sensible way to preserve global structure while adding high-resolution texture. The qualitative figures show outputs that do look SAR-like, with plausible speckle and detail. The authors also acknowledge that the pretrained ControlNet modules were designed for optical images and that this mismatch "likely impacted" performance.

Now the soft spots. The central claim is that the pipeline "effectively transforms" spaceborne or simulated SAR into airborne SAR representations. The evidence is purely qualitative. No quantitative metrics, no comparison against real airborne SETHI images of the same scenes, no ablation (e.g., a no-ControlNet or simple-upscaler baseline), no downstream task validation. The paper's own discussion undercuts the claim by admitting the ControlNet conditional signal is poorly matched to SAR. So you can't tell whether the outputs are faithful domain adaptation or just plausible texture generation from the prior. That gap is load-bearing, not a minor omission. The dataset is private, so others can't verify the training data or replicate the results. These are addressable, but as it stands the paper is a proof-of-concept, not a validated method.

To me, the reader's score of 3 on soundness is about right. But the significance and novelty are not inflated: for the SAR community, a low-cost way to generate high-resolution training data from satellite or simulated inputs would be genuinely useful, and this is the first SDXL+ControlNet application I know of in that space.

I'd send this to peer review, but with the clear expectation of major revision. The authors should add quantitative evaluation (e.g., distribution similarity metrics, edge fidelity, comparisons to real airborne acquisitions), an ablation of the ControlNet strength, and ideally release a small benchmark or at least a few aligned input/output pairs. Who benefits? Remote sensing researchers working on SAR data augmentation or super-resolution. I wouldn't cite it in my own work, but it's a fair candidate for a workshop or a short paper once strengthened.

Recommendation: engage with it, but only conditionally — request metrics and comparisons before accepting.

Referee Report

4 major / 5 minor

Summary. The paper proposes a multi-resolution synthesis pipeline that uses Stable Diffusion XL (SDXL), fine-tuned with LoRA on a 110K-image dataset of ONERA SETHI X-band airborne SAR data, together with two ControlNet modules (Canny and Tile) to transform TerraSAR-X satellite images and EMPRISE simulated images into 40 cm resolution 'airborne SAR representations.' The method is presented in Sections II and III, with results shown only as qualitative side-by-side figures (Figs. 2 and 3). The authors claim novelty and usefulness for data augmentation, while also acknowledging in Section IV that the pre-trained ControlNet modules are mismatched to SAR imagery and that this 'likely impacted the overall performance.'

Significance. If the central claim were demonstrated, the pipeline would provide a practical data-augmentation tool for high-resolution SAR applications, and the construction of a 110K paired SAR-optical dataset from ONERA archives is a nontrivial asset. The paper is also transparent about its main limitation. However, the manuscript provides no quantitative evidence that the outputs are faithful airborne SAR representations rather than plausible SAR-like textures: there are no fidelity metrics, no comparison to real SETHI acquisitions of the same scene, no ablations, and no downstream-task validation. Because the claimed contribution is empirical, the current evidence is insufficient to assess the method's validity or practical value.

major comments (4)

[Section III-B and Figures 2–3] The central claim that satellite or simulated SAR images are transformed into airborne SAR representations is supported only by qualitative side-by-side examples. No quantitative metric is reported: there is no comparison against real SETHI airborne images of the same scene, no structure-preservation measure (e.g., edge or contour correlation between input and output), no radiometric or speckle statistics, and no error bars. Under the alternative hypothesis that the diffusion prior generates plausible-looking SAR texture regardless of input content, the figures would appear similar, so the evidence does not distinguish faithful domain adaptation from hallucinated texture. The paper needs quantitative evaluation before the central claim can be assessed.
[Section IV] The authors concede that the pre-trained ControlNet modules were optimized for optical images and that 'this mismatch complicates the ability ... to effectively condition SAR images' and 'likely impacted the overall performance,' and they also state that 'significant modifications to the original structure' can occur. This directly undermines the assumption that scene geometry is preserved, which is load-bearing for a transformation method. An ablation comparing SDXL+LoRA without ControlNet, SDXL+LoRA with optical ControlNet, and a simple upscaler, along with a quantitative fidelity measure, is needed to determine what the ControlNet modules contribute and whether the input structure is retained.
[Sections II-C and III-B] Reproducibility is insufficient. The LoRA fine-tuning hyperparameters (rank, learning rate, number of steps) and the per-resolution training-set composition are not specified; the dataset splitting, captioning procedure, and filtering criteria are described only qualitatively; and the ControlNet strength and end-percent values are given without sensitivity analysis. Without these details, the reader cannot reproduce the experiments or judge whether the selected parameters are principled rather than cherry-picked.
[Section V] The conclusion frames the pipeline as generating useful data for SAR applications, but no downstream-task experiment is reported. If the intended use is data augmentation, a detection or segmentation experiment trained on the synthetic data and evaluated on real airborne SAR would provide a concrete, falsifiable test of utility. The absence of such an experiment leaves the application claim unsupported.

minor comments (5)

[Abstract and Section I] There are several typos and grammatical errors, for example 'airborn' in the abstract, 'airbone' in Section I, and 'creativly' in Section I; these should be corrected.
[Figures 2 and 3] The figures lack scale bars, coordinate information, and an explicit statement that the top and bottom rows show the same geographical scene; this information is essential for evaluating whether structure is preserved.
[Equation (2)] The notation for the product term using 'Q' is nonstandard and ambiguous; it should be written with a product symbol, and the definitions of alpha_t and beta_t should be stated more carefully.
[References] Reference [2] is incomplete (no title, venue, or year), and the '3.5 billion parameters' claim for SDXL should be attributed to the cited Podell et al. paper at its first mention rather than to reference [1] in the abstract.
[Section I] The claim 'to the best of our knowledge, we are the first to introduce this approach in the literature' requires a proper related-work comparison; without it, the novelty claim cannot be verified.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the pipeline is an empirical generative system, and its self-citations are contextual rather than load-bearing.

full rationale

The paper does not claim to derive a quantitative prediction from first principles; it presents an empirical image-to-image pipeline. The central claim is that a fine-tuned Stable Diffusion XL with ControlNet conditioning can transform TerraSAR-X or simulated SAR images into airborne-like SAR representations. This claim is supported by qualitative figures and not by an independent external metric, but the absence of a quantitative ground-truth comparison is an evidential gap, not a circular reduction. No equation in the paper defines the output in terms of the fitted target, and no fitted parameter is renamed as a prediction. The self-citations [6] and [7] are used only to describe prior related work by the same group ('Lately, we [7] presented a conditional multi-resolution latent diffusion pipeline...'), and the current paper does not rely on them as the justification for its own validity. The Section IV admission that pre-trained ControlNet modules were 'originally fine-tuned for controlling optical images' and that this mismatch 'likely impacted the overall performance' is a limitation statement, not a hidden input smuggled in through citation. The training data and the evaluation domain are indeed both ONERA SETHI airborne SAR imagery, which is a methodological concern about how 'realism' is judged, but evaluating a generative model by comparing its outputs to the distribution it was trained on is the standard construction of the task, not a self-definitional fallacy. Therefore no specific circular step can be exhibited, and the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities. It relies on several domain assumptions about transferring optical-domain generative models to SAR, the most fragile being the optical-trained ControlNet conditioning. The manually chosen ControlNet strengths and end percentages, plus unreported LoRA hyperparameters, are free parameters that limit reproducibility.

free parameters (4)

ControlNet strength (both modules) = 0.8
Manual setting of the influence of ControlNet conditioning in both pipeline stages; chosen by the authors without reported optimization or sensitivity analysis.
ControlNet end percent, Module 1 (Canny and Tile) = 0.7 and 0.8
User-set parameters in ComfyUI that stop ControlNet conditioning early in the denoising process; chosen by hand.
ControlNet end percent, Module 2 (Canny and Tile) = 0.8 (both)
Same as above for the second upscaling stage.
LoRA fine-tuning hyperparameters (rank, learning rate, steps) = not stated
The paper does not report any LoRA training details, so these are unknown free parameters that affect the fine-tuned SD80 and SD40 models.

assumptions (4)

domain assumption Pre-trained Stable Diffusion XL prior transfers to the SAR image domain via LoRA fine-tuning.
The paper assumes a diffusion model trained on optical images can be fine-tuned to generate realistic SAR textures; only qualitative evidence is provided.
domain assumption ControlNet modules trained on optical images provide useful spatial conditioning for SAR images.
Central to the pipeline's structure; the paper itself notes this assumption is questionable, stating in Section IV that it 'likely impacted the overall performance.'
domain assumption Optical captions generated by CogVLM2 describe the content of SAR images.
Used for text conditioning, but SAR and optical images differ radically in appearance; no validation is provided that the captions are semantically correct for SAR content.
domain assumption Latent upscaling from 512 to 1024 to 2048 preserves the global geometric structure of the scene.
The pipeline assumes upscaling and refinement steps do not distort the scene; the paper acknowledges that ControlNet strength can cause 'significant modifications to the original structure.'

how reviews work

0 comments

Cite this review

Pith. "Pith review of From Spaceborne to Airborne: SAR Image Synthesis Using Foundation Models for Multi-Scale Adaptation." pith.science (2026). https://pith.science/paper/XZ6NFV3F

@misc{pith2026250503844,
  author       = {Pith},
  title        = {Pith review of: From Spaceborne to Airborne: SAR Image Synthesis Using Foundation Models for Multi-Scale Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XZ6NFV3F}},
  note         = {Machine review of arXiv:2505.03844}
}

read the original abstract

The availability of Synthetic Aperture Radar (SAR) satellite imagery has increased considerably in recent years, with datasets commercially available. However, the acquisition of high-resolution SAR images in airborne configurations, remains costly and limited. Thus, the lack of open source, well-labeled, or easily exploitable SAR text-image datasets is a barrier to the use of existing foundation models in remote sensing applications. In this context, synthetic image generation is a promising solution to augment this scarce data, enabling a broader range of applications. Leveraging over 15 years of ONERA's extensive archival airborn data from acquisition campaigns, we created a comprehensive training dataset of 110 thousands SAR images to exploit a 3.5 billion parameters pre-trained latent diffusion model \cite{Baqu2019SethiR}. In this work, we present a novel approach utilizing spatial conditioning techniques within a foundation model to transform satellite SAR imagery into airborne SAR representations. Additionally, we demonstrate that our pipeline is effective for bridging the realism of simulated images generated by ONERA's physics-based simulator EMPRISE \cite{empriseem_ai_images}. Our method explores a key application of AI in advancing SAR imaging technology. To the best of our knowledge, we are the first to introduce this approach in the literature.

Figures

Figures reproduced from arXiv: 2505.03844 by the authors.

**Figure 1.** Creative upscaling pipeline an effective ground resolution of 80 cm. This intermediate output is subsequently upscaled to 2048×2048 and processed by ControlNet (Module 2), which also uses the SDXL Canny and SDXL Tile filters, both set with an end percent of 0.8 and a strength of 0.8. This step refines the image through the SD40 model (Stable Diffusion XL fine-tuned with a LoRA at a 40cm resolution) to achieve a fina… view at source ↗

**Figure 2.** Enhanced TerraSAR-X satellite images to aerial resolution Top row: TerraSAR-X images (512x512, 80cm). Bottom [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Enhanced simulated images from ONERA’s Radar simulator with creative content Top row: simulated images (512x512, [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

Quantitative Comparison of Fine-Tuning Techniques for Pretrained Latent Diffusion Models in the Generation of Unseen SAR Images
cs.CV 2025-06 conditional novelty 6.0 of 10

A systematic comparison of fine-tuning strategies for Stable Diffusion XL on 100k real SAR images finds that full UNet fine-tuning with LoRA text encoders and a learned <SAR> token gives the best generation quality.

Reference graph

Works this paper leans on

10 extracted references · 5 canonical work pages · cited by 1 Pith paper

[1]

R \'e mi Baqu \'e , Philippe Dreuillet, and H \'e l \`e ne M. Oriot. Sethi : Review of 10 years of development and experimentation of the remote sensing platform. 2019 International Radar Conference (RADAR) , 2019

work page 2019
[2]

Examples of ai generated sar images, 2024

DEMR-SEM ONERA. Examples of ai generated sar images, 2024

work page 2024
[3]

Denoising diffusion probabilistic models, 2020

Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020

2020
[4]

Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio

Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks, 2014

2014
[5]

High-resolution image synthesis with latent diffusion models, 2022

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022

2022
[6]

Sar image synthesis using text conditioned pre-trained generative ai models

Nicolas Trouve, Nathan Letheule, Olivier Leveque, Ilias Rami, and Elise Colin. Sar image synthesis using text conditioned pre-trained generative ai models. In Proceedings of EUSAR 2024; 15th European Conference on Synthetic Aperture Radar , Munich, Germany, 2024. VDE, VDE,ITG

work page 2024
[7]

Synthesizing sar images with generative ai: Expanding to large-scale imagery

Debuysere Solene, Trouve Nicolas, Letheule Nathan, Leveque Olivier, and Colin Elise. Synthesizing sar images with generative ai: Expanding to large-scale imagery. In Proceedings of RADAR 2024 , Rennes, France, 2024

work page 2024
[8]

Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023

Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023

2023

Show all 10 references

[9]

Adding conditional control to text-to-image diffusion models, 2023

Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023

2023
[10]

[ h ): /`ވ&3|.nVU.. >aD@ i

11em plus .33em minus .07em 4000 4000 100 4000 4000 500 `\.=1000 = #1 \@IEEEnotcompsoconly \@IEEEcompsoconly #1 * [1] 0pt [0pt][0pt] #1 * [1] 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEcompsocnotconfonly \@IEEEauthorblockAstyle \@IEEEcompsocnotconfonly \@IEEEco...

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [1]

R \'e mi Baqu \'e , Philippe Dreuillet, and H \'e l \`e ne M. Oriot. Sethi : Review of 10 years of development and experimentation of the remote sensing platform. 2019 International Radar Conference (RADAR) , 2019

work page 2019

[2] [2]

Examples of ai generated sar images, 2024

DEMR-SEM ONERA. Examples of ai generated sar images, 2024

work page 2024

[3] [3]

Denoising diffusion probabilistic models, 2020

Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020

2020

[4] [4]

Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio

Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks, 2014

2014

[5] [5]

High-resolution image synthesis with latent diffusion models, 2022

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022

2022

[6] [6]

Sar image synthesis using text conditioned pre-trained generative ai models

Nicolas Trouve, Nathan Letheule, Olivier Leveque, Ilias Rami, and Elise Colin. Sar image synthesis using text conditioned pre-trained generative ai models. In Proceedings of EUSAR 2024; 15th European Conference on Synthetic Aperture Radar , Munich, Germany, 2024. VDE, VDE,ITG

work page 2024

[7] [7]

Synthesizing sar images with generative ai: Expanding to large-scale imagery

Debuysere Solene, Trouve Nicolas, Letheule Nathan, Leveque Olivier, and Colin Elise. Synthesizing sar images with generative ai: Expanding to large-scale imagery. In Proceedings of RADAR 2024 , Rennes, France, 2024

work page 2024

[8] [8]

Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023

Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023

2023