Pith. sign in

REVIEW 4 major objections 5 minor 14 references

PriSAR: 3D Geometric-Prior-Guided Diffusion for Parameter-Controlled SAR Image Generation

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a lightweight 3D point-cloud prior, derived by ray casting on 3D models, guides a diffusion model to generate SAR images that closely track the specified azimuth angle, sharply improving viewpoint consistency over text

desk verdict Useful engineering idea, but the causal claim for the geometric prior is confounded by the baseline and the paper has serious version-control issues. read the letter →

arxiv 2607.22963 v1 pith:X6XQVTT6 submitted 2026-07-25 eess.IV cs.CV

Fan Zhang , Xuanting Wu , Fei Ma , Qiang Yin , Yuxin Hu This is my paper · ORCID
classification eess.IVcs.CV
keywords syntheticapertureradarimagegenerationdiffusionmodelsgeometricpriorraycastingpointclouddataaugmentationazimuthcontrol
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

PriSAR tries to solve the problem of controllable synthetic aperture radar (SAR) image generation when real data are sparse. Its central proposition is that explicit geometry, in the form of simulated scattering-center point clouds, is what text-conditioned diffusion models lack when they produce structurally inconsistent targets. The paper builds a lightweight ray-casting prior that approximates multi-bounce scattering, fuses it with text and image features through a FiLM-based network, and adapts Stable Diffusion 3.5 Medium via LoRA. On a real four-aircraft dataset, the method reports SSIM 0.812 and azimuth classification F1 0.940, versus 0.738 and 0.782 for the text-only baseline, and the gain is largest on the angle-estimation task. The authors deliberately frame the prior as generation guidance, not electromagnetic reconstruction.

What carries the argument

The operative mechanism is the viewpoint-aware point-cloud prior built by recursive ray casting on a 3D mesh (Eqs. 4-7 with the Pseudo-RCS response function Ψ). Each ray carries position, direction, and energy; upon hitting the mesh, energy is attenuated by a heuristic geometric response and re-emitted by specular-plus-diffuse reflection, yielding single-, double-, and triple-bounce scattering points colored by energy and bounce order. This point cloud is encoded by a Point Transformer V3, projected to the slant-range plane, and fused with T5/CLIP text features and a training-time image anchor through a FiLM-based fusion network, while LoRA adapts the SD3.5 Medium backbone. The prior's job i

What would settle it

Train the identical pipeline with a control prior: replace the ray-cast point cloud with a fixed or random point cloud (same encoder, same fusion, same LoRA). If SSIM and azimuth F1 do not drop substantially, the geometric content of the prior is not what carries the gain. Additionally, extract actual scattering centers from the real SAR images (e.g., by sparse reconstruction) and compare them against the simulated point clouds; if the simulated point clouds do not align with real scattering-center locations at corresponding azimuths, the physical interpretability of the prior is in question.

Watch

Extended reading notes

Core claim

The central claim is that a 3D-model-derived geometric prior, encoded as a view-dependent point cloud with per-point energy and reflection order, can be injected into a latent diffusion model to enforce observation-geometry consistency in SAR synthesis. The point cloud is produced by a recursive ray-tracing model with hand-set gains for edge, orientation, and corner-reflector-like structures (the Pseudo-RCS response function). Combined with a five-stage multi-modal fusion network (projection normalization, gated fusion, FiLM modulation, refinement, cosine-similarity anchoring) and LoRA fine-tuning of the SD3.5 Medium backbone, this prior yields generated images whose azimuth-dependent struct

Load-bearing premise

The load-bearing premise is that the hand-designed Pseudo-RCS ray-casting model, whose gains were manually tuned on the same SSIM and azimuth metrics used for evaluation, approximates real SAR scattering-center geometry closely enough that the measured improvements come from geometric guidance rather than from the tuning process or the extra conditioning pathway.

Editorial extensions

If this is right

  • If the geometric prior is the cause of the gain, sparse-angle SAR datasets can be densified with synthetic views at intermediate azimuths, improving downstream classifiers, as the multi-label experiment suggests.
  • The method establishes a tractable middle ground between physics-heavy electromagnetic simulation and purely data-driven generation, with prior construction costing 48.3 s per azimuth on CPU and inference only about 2 s more than the backbone.
  • Viewpoint consistency becomes a controllable attribute of generation, so azimuth, polarization, and class can be specified independently in the prompt and honored in the image.
  • The design suggests geometric priors matter most for fine-grained orientation (azimuth) rather than coarse class identity, where text-only models already saturate near 0.99 F1.

Reading between the lines

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

  • The F1 gain on azimuth (0.782 to 0.940) could be partly inherited from the metric: if generated images collapse toward a canonical pose or average shape, a classifier can latch onto a stable artifact; the t-SNE separation may reflect a simplified manifold, not true pose fidelity. A direct test would be to measure the distribution of predicted azimuths against the requested azimuths, not just class
  • The point-cloud prior is not validated against real scattering centers. A natural next step is to compare simulated point clouds with scattering-center maps extracted from real SAR imagery (e.g., via CLEAN or sparse reconstruction) and to retune the hand-set gains against that data.
  • The method's dependence on a 3D model per category limits deployment to targets with available CAD meshes; if the prior is what carries viewpoint control, the approach may not transfer to unknown target classes without a geometry source.
  • The abstract claims the same protocol works on five MSTAR vehicle classes, but the main body omits those results; a reproducible public-benchmark comparison would strengthen the claim.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes PriSAR, a diffusion-based framework for SAR image generation with explicit control over observation geometry. A 3D model of the target is converted into a viewpoint-dependent point cloud by a lightweight ray-casting procedure with multiple bounces and a hand-designed pseudo-RCS response; this point cloud is encoded with Point Transformer V3 and fused with text and image features through a five-stage FiLM-based fusion network. Stable Diffusion 3.5 Medium is adapted with LoRA, and inference uses text plus geometry conditioning. The method is evaluated on a self-collected four-class aircraft SAR dataset under a sparse-angle protocol, with quantitative comparisons against GAN, autoregressive, and diffusion baselines. Reported results include PSNR 31.36, SSIM 0.812, and azimuth-classification F1 0.940, versus 25.23, 0.738, and 0.782 for the SD3.5m baseline. The paper also reports downstream classification gains, computational cost, and parameter-sensitivity checks. The central claim is that the 3D geometric prior improves structural fidelity and viewpoint consistency for controllable SAR generation.

Significance. If the central claim is established, the paper would make a useful contribution: it proposes a computationally tractable middle ground between full electromagnetic simulation and purely data-driven SAR generation, and it addresses a real limitation of text-to-image models—weak geometric control under varying azimuth. The point-cloud prior is lightweight, the use of LoRA is appropriate for small SAR datasets, and the downstream data-augmentation experiment is a sensible application-oriented evaluation. The paper is also honest about the intended scope: the prior is a guidance mechanism rather than an electromagnetic reconstruction. However, the evidence for the central causal claim is currently incomplete: the main baseline is not a controlled ablation of the geometric prior, the heuristic parameters were tuned on the same metrics used for the final comparison, and the azimuth-consistency metric is a downstream classifier accuracy rather than a direct geometric-adherence measurement. These issues are fixable with additional experiments and reporting, so the core idea is worth pursuing, but the current support is not yet sufficient.

major comments (4)
  1. [§3.3.2, Tables 3 and 4] The comparison that carries the central claim is between PriSAR and 'SD3.5m', described as the backbone 'without 3D point-cloud features'. But this baseline also omits the five-stage multi-modal fusion network of §2.3 and the training-time image-feature branch. The observed gains in SSIM and azimuth F1 could therefore be caused by the added fusion capacity or the auxiliary image supervision rather than by the point-cloud prior itself. A controlled ablation is needed: the same fusion network and training protocol, with and without the point-cloud branch (and ideally with a non-geometric control condition, e.g., a fixed random point cloud), to isolate the contribution of the geometric prior.
  2. [§2.2, Table 1] The default heuristic parameters in Table 1 (Kmax, ζ, α_edge, α_vert, α_struct, thresholds, noise scales) are stated to have been manually selected and tuned in preliminary checks using SSIM and azimuth-related consistency—the same metrics used for the final comparison in Tables 3 and 4. With no separate validation split or multi-seed variance reported, part of the observed improvement may reflect selection on the evaluation metric. The authors should report tuning on a held-out validation set and present final test metrics with multiple random seeds.
  3. [§3.4.4, Table 4] Azimuth 'consistency' is operationalized as the F1 of a classifier trained on a dataset that includes generated images. Those generated images were conditioned on the exact azimuth labels that the classifier is asked to predict. While the test set is real and held out, the metric can partly measure label propagation from the conditioning signal rather than an independent geometric property of the generated images. The paper should report a more direct geometric-adherence metric (e.g., comparison of generated bright-spot locations with ray-cast point-cloud projections) and should also report variance across classifier training runs.
  4. [§2.2, Eqs. (4)-(7)] The Pseudo-RCS response function Ψ is a hand-designed heuristic with several free gains and thresholds, but the paper does not validate the simulated point clouds against real SAR scattering-center distributions or against measured SAR images. Since the geometric prior is the proposed cause of the improvement, some form of validation is needed—for example, comparing the projected point clouds to real SAR bright-spot locations, or ablating Ψ by replacing it with uniform energy weighting while keeping the same ray-casting geometry. This would also address whether the observed gains come from the geometric prior per se or from the specific hand-tuned response model.
minor comments (5)
  1. [Abstract and Title] The abstract uses the name 'GeoDiff-SAR', while the title and main text use 'PriSAR'. Please make the naming consistent throughout.
  2. [References] The reference for Point Transformer V3 has a typo: 'Wu, X.,aa Jiang, L.' should be 'Wu, X., Jiang, L.'
  3. [§3.3.1] The paired evaluation protocol is described only briefly. Please clarify how the 'semantic labels from the held-out test set' are converted into text prompts and whether the same prompt is used for all models. Also state the number of evaluation samples.
  4. [§3.4.1] The downstream experiment uses a sparse 10° training split and generates 5° intermediate views. It would be helpful to report the exact number of generated images per class/polarization/azimuth and the real/generated ratio in the mixed training set, since this affects the interpretation of the downstream gains.
  5. [§3.3.2, Table 3] No error bars or standard deviations are reported for any of the metrics in Table 3. For generative models, even small seed changes can alter PSNR/SSIM noticeably; please report means and variances over at least three seeds.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity: the geometric-prior heuristic parameters were tuned on the same SSIM/azimuth-consistency metrics later reported as evidence; otherwise the pipeline is largely self-contained with external held-out test data.

  1. fitted input called prediction [Section 2.2, Construction of Geometric Priors (paragraph after Eq. 7, before Table 1 discussion)]
    "In preliminary sensitivity checks, increasing Kmax from 1 to 3 and increasing ζ from 0 to 0.8 both improved SSIM and azimuth-related consistency, and the best setting was obtained at Kmax = 3 and ζ = 0.8, which is therefore used throughout this paper."

    The two key heuristic parameters of the pseudo-RCS prior, maximum bounce count Kmax and diffuse coefficient ζ, are explicitly selected by optimizing the exact quantities later used to validate the prior: SSIM in Table 3 and azimuth-related consistency in Table 4. The paper then reports PriSAR's higher SSIM (0.812 vs 0.738) and azimuth F1 (0.939 vs 0.782) as evidence that the 3D geometric prior improves viewpoint adherence. Since the prior was tuned to maximize those same metrics, part of the observed gain is a selection effect rather than an independent confirmation of the geometric prior. The point-cloud model is also never validated against real scattering centers, so the causal role of geometry is not separately identified.

full rationale

The paper does not contain a self-definitional derivation: the ray-casting Pseudo-RCS model, the fusion network, and the diffusion objective are not defined in terms of the reported outputs. The self-citations (e.g., Ma et al. 2022a,b; Zhang et al. 2019) are background references on SAR segmentation and GAN-based augmentation and are not load-bearing for the central claim. No uniqueness theorem is imported from the authors' prior work. The downstream classification evaluation is on a held-out real SAR test set, so the augmentation claim is not circular by construction. The main circularity concern is the manual tuning of Kmax and ζ on the same SSIM/azimuth-consistency metrics that are later used as evidence; this is a real but partial selection effect, since the geometric prior still has substantial independent content and the final test set is external. A further experimental limitation, not scored as circularity, is that the SD3.5m baseline lacks the fusion network and training-time image branch, so the geometric prior is not fully isolated; this is a correctness/ablation concern rather than a definitional reduction.

Assumptions & free parameters 12 free parameters · 5 assumptions · 0 invented entities

The central claim relies on a set of hand-set heuristic parameters in the point-cloud simulator and several domain assumptions about the fidelity of 3D models, the pseudo-RCS scattering proxy, and the transferability of a natural-image diffusion backbone to SAR. These are the main unverified inputs the reader is asked to accept.

free parameters (12)
  • Kmax = 3
    Maximum number of reflections; Section 2.2 reports it was chosen after sensitivity checks on SSIM/azimuth metrics.
  • zeta = 0.8
    Diffuse scattering coefficient; tuned to 0.8 in sensitivity checks per Section 2.2.
  • alpha_edge = 2.0
    Edge enhancement factor in the Pseudo-RCS response function; manual heuristic in Table 1.
  • alpha_vert = 2.0
    Vertical-surface gain in the Pseudo-RCS response function; manual heuristic in Table 1.
  • alpha_struct = 1.6
    Structural/dihedral scattering gain; manual heuristic in Table 1.
  • tau_area = 0.01
    Small-face threshold for edge enhancement; manual heuristic in Table 1.
  • tau_vert = 0.3
    Verticality threshold for orientation gain; manual heuristic in Table 1.
  • mu = 0.03
    Absorption coefficient in the energy-update equation; manual heuristic in Table 1.
  • tau_min = 0.03
    Reflection termination threshold for point-cloud construction; manual heuristic in Table 1.
  • sigma_scan = 0.3
    Initial ray perturbation scale; manual heuristic in Table 1.
  • sigma_reflect = 0.1
    Reflection noise scale; manual heuristic in Table 1.
  • lambda = 0.5
    Residual scaling factor in the FiLM modulation (Eq. 13), chosen to stabilize training.
assumptions (5)
  • ad hoc to paper The Pseudo-RCS response function Psi (Eqs. 4-5) with heuristic gains alpha_edge, alpha_vert, alpha_struct and thresholds is a sufficient proxy for SAR backscattering geometry.
    Introduced in Section 2.2; never validated against measured scattering centers or EM simulation, yet it is central to the geometric prior.
  • domain assumption The 3D mesh models of the four aircraft represent the real field-collected targets closely enough for geometry guidance.
    Dataset construction in Section 3.1.1; no fidelity check between CAD models and the real aircraft is reported.
  • domain assumption Projecting the point-cloud feature onto the slant-range plane preserves the geometric information needed by the 2D diffusion model.
    Section 2.2 describes the adaptive projection mapping; no ablation tests alternative projections.
  • domain assumption A diffusion model pre-trained on natural images, adapted via LoRA, can learn SAR-specific texture and clutter from roughly 1,870 images.
    Section 2.4 relies on this transfer-learning assumption without independent verification.
  • domain assumption The T5/CLIP text encoders can represent azimuth angle and polarization as meaningful conditions from the templated prompt.
    Section 2.3 uses the text condition as a baseline control; if encoders cannot resolve fine azimuth differences, the text-only baseline is handicapped.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PriSAR: 3D Geometric-Prior-Guided Diffusion for Parameter-Controlled SAR Image Generation." pith.science (2026). https://pith.science/paper/X6XQVTT6

@misc{pith2026260722963,
  author       = {Pith},
  title        = {Pith review of: PriSAR: 3D Geometric-Prior-Guided Diffusion for Parameter-Controlled SAR Image Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X6XQVTT6}},
  note         = {Machine review of arXiv:2607.22963}
}
read the original abstract

Synthetic aperture radar (SAR) image generation can mitigate data scarcity, but controllablegeneration under sparse observation angles remains difficult. Recent SAR generative studies im-prove texture realism, yet explicit geometry-aware control is still limited. This paper studiesthe focused and verifiable setting of intermediate-azimuth completion: 3D-model-derived geo-metric priors guide a diffusion model to synthesize the views missing from sparse-angle trainingdata. GeoDiff-SAR constructs a lightweight multi-bounce ray-tracing prior, encodes the result-ing point cloud, and fuses it with text conditioning while adapting Stable Diffusion 3.5 Mediumthrough low-rank adaptation. On a real four-category aircraft dataset, GeoDiff-SAR reaches anSSIM of 0.812 and azimuth consistency of 0.940, compared with 0.738 and 0.782 for the text-conditioned SD3.5 Medium baseline. The same sparse-angle protocol on five MSTAR vehicleclasses yields an SSIM of 0.878 and azimuth consistency of 0.917. These results support theconclusion that a lightweight 3D geometric prior improves viewpoint adherence for controllableSAR generation; it is intended as generation guidance rather than high-fidelity electromagneticreconstruction.

Figures

Figures reproduced from arXiv: 2607.22963 by the authors.

Figure 1
Figure 1. Conceptual radar-chart comparison of five representative SAR image generation paradigms. The axes denote generation speed, azimuth control, physical interpretability, parameter simplicity, and geometric structure preservation. The chart is intended to illustrate the relative positioning of different paradigms; PriSAR occupies an intermediate position between rigorous simulation and purely data-driven generation by c… view at source ↗
Figure 2
Figure 2. Illustration of the research motivation and paradigm comparison. (Left) Traditional electromagnetic simulations offer strong physical interpretability but are computationally expensive and require detailed parameter settings (e.g., antenna dimensions and PRF). (Middle) Generic text-to-image models provide efficient generation but often lack explicit geometric control, which may lead to structural inconsistency. (Rig… view at source ↗
Figure 3
Figure 3. Overall architecture of PriSAR. (a) Construction of geometric priors: 3D models are processed by a point￾cloud module to extract view-dependent scattering cues, such as multi-bounce reflections and occlusion structures, which serve as geometric guidance. (b) Multi-modal fusion network: these features are combined with textual and visual conditions to guide the Stable Diffusion 3.5 Medium backbone, which is adapted t… view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: In this geometric configuration, we define a specific coordinate system aligned with the target’s orientation, where the X-axis is designated as West (0 ◦ ) and the Y -axis as South (90◦ ), with the azimuth angle increasing in a counter￾clockwise direction. Based on th…
Figure 4
Figure 4. Figure 4: Schematic diagram of object-centric radar observation geometry for point cloud simulation. The coordinate system sets the X-axis to West (0 ◦ ) and the Y -axis to South (90◦ ). The radar position is defined by azimuth ϕ and depression ψ, forming the LOS vector toward t…
Figure 5
Figure 5. Figure 5: Schematic diagram of radar illumination at different azimuth angles and the corresponding calculated point cloud models. The point clouds are color-coded by reflection type (blue for single, yellow for double, red for triple, and purple for multiple bounces), effective…
Figure 6
Figure 6. Figure 6: The overall architecture of the proposed multi-modal fusion network. It comprises five cascaded modules: (1) Projection Normalization aligns multi-modal features into a unified high-dimensional space; (2) Gated Fusion dynamically assigns weights to text, image, and poi…
Figure 7
Figure 7. Figure 7: The schematic diagram of the LoRA working principle applied in the SD3.5 backbone. The training objective follows the standard noise-prediction formulation of latent diffusion models. Given a real SAR image x ∈ Dtrain, we first use the VAE encoder to map it into the la…
Figure 8
Figure 8. Figure 8: Qualitative generation results of PriSAR across different aircraft categories and azimuth angles. The figure presents side-by-side comparisons of real SAR images and generated samples for King Air 350i (55◦ ), Cessna 208 (45◦ ), Pilatus PC 12 (220◦ ), and Kodiak 100 (2…
Figure 9
Figure 9. Figure 9: Data structure and composition strategy. The figure illustrates the three key components of our experimental data: (1) Real SAR Images Dataset for Generative Models Training: A sparse subset of real data sampled at coarse 10◦ azimuth intervals (e.g., 0 ◦ , 10◦ ), used …
Figure 10
Figure 10. Figure 10: Visual quality comparison between PriSAR and representative baseline models. Each column corresponds to a distinct configuration: column 1 shows Kodiak 100 (90◦ , HH); column 2 shows King Air 350i (55◦ , HV); column 3 shows Cessna 208 (45◦ , VH); and column 4 shows Pi…
Figure 11
Figure 11. Figure 11: Visualization of the multi-modal correspondence between text, point clouds, and generated images. (a) Point cloud generation corresponding to azimuth angles: The intermediate 3D point cloud priors generated under varying observation geometries (e.g., 0 ◦ to 315◦ ), se…
Figure 12
Figure 12. Figure 12: Workflow of the downstream evaluation. Sparse 10◦ real data are first used to train the generative models; PriSAR then synthesizes the missing intermediate azimuths with 3D-guided priors, and the resulting mixed dataset is used for downstream classification. During da…
Figure 13
Figure 13. Figure 13: Visualization of downstream classification results: (a) azimuth confusion curves, (b) relative polarization-class gains over the real-data baseline, and (c) relative aircraft-type gains. PriSAR appears to preserve more discriminative cues than the compared baselines. …
Figure 14
Figure 14. Figure 14: T-SNE visualization of feature distributions in the azimuth estimation task. PriSAR forms clearer viewpoint￾related clusters than the compared baselines. To conduct a fine-grained ablation study on the efficacy of the proposed 3D geometric prior, we quantitatively eva…
Figure 15
Figure 15. Figure 15: Fine-grained ablation over the full 360◦ azimuth range. PriSAR exhibits smoother Precision, Recall, and F1 trends, indicating stronger viewpoint consistency than the real-data baseline and vanilla SD3.5m. As illustrated, the Real Image baseline forms a compact inner c…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 2 canonical work pages

  1. [4]

    IEEE Trans

    General sparse adversarial attack method for SAR images based on keypoints. IEEE Trans. Aerosp. Electron. Syst. 61, 14943–14960. doi: 10.1109/TAES.2025.3588821. Goodfellow, I., Pouget-Abadie, J., Mirza, M., et al.,

    arXiv 2025

  2. [5]

    IEEE Trans

    X-Fake: Juggling utility evaluation and explanation of simulated SAR images. IEEE Trans. Image Process. 34, 7830–7844. doi: 10.1109/TIP.2025.3634988. Huang, Z., Zhang, X., Tang, Z., Xu, F., Datcu, M., Han, J.,

    arXiv 2025

  3. [6]

    IEEE Geosci

    Generative artificial intelligence meets synthetic aperture radar: A survey. IEEE Geosci. Remote Sens. Mag., early access. doi: 10.1109/MGRS.2024.3483459. Isola, P., Zhu, J.-Y., Zhou, T., et al.,

    arXiv 2024

  4. [9]

    FiLM: Visual reasoning with a general conditioning layer. Proc. AAAI Conf. Artif. Intell. 32, 3942–3951. doi: 10.1609/aaai.v32i1.11671. Podell, D., English, Z., Lacey, K., et al.,

    doi:10.1609/aaai.v32i1.11671

  5. [11]

    IEEE Trans

    A low-cost polarimetric radar system based on mechanical rotation and its signal processing. IEEE Trans. Aerosp. Electron. Syst. 61, 4744–4765. doi: 10.1109/TAES.2024.3507776. Song, J., Meng, C., Ermon, S.,

    arXiv 2024

  6. [12]

    IEEE Trans

    CRS-Diff: Controllable remote sensing image generation with diffusion model. IEEE Trans. Geosci. Remote Sens. 62, 1–14. doi: 10.1109/TGRS.2024.3453414. Van den Oord, A., Vinyals, O., Kavukcuoglu, K.,

    arXiv 2024

  7. [14]

    IEEE Geosci

    Scattering characteristics guided network for ISAR space target component segmentation. IEEE Geosci. Remote Sens. Lett. 22, Art. no. 4009505. doi: 10.1109/LGRS.2025.3576662. Zhu, J.-Y., Park, T., Isola, P., et al.,

    arXiv 2025

  8. [1374]

    Schreiber, E., Gupta, I.J.,

    doi: 10.3390/RS11111374. Schreiber, E., Gupta, I.J.,

    work page doi:10.3390/rs11111374

Show all 14 references
  1. [2015]

    Deep convolutional neural networks for ATR from SAR imagery. Proc. SPIE 9475, Algorithms for Synthetic Aperture Radar Imagery XXII, 94750F. doi: 10.1117/12.2176558. Mou, C., Wang, X., Xie, L., et al.,

    doi

  2. [2018]

    Generative adversarial networks for SAR image realism. Proc. SPIE 10647, Algorithms for Synthetic Aperture Radar Imagery XXV, 1064709. Ma, F., Zhang, F., Xiang, D., Yin, Q., Zhou, Y., 2022a. Fast task-specific region merging for SAR image segmentation. IEEE Trans. Geosci. Remo...

    2022

  3. [2020]

    PyTorch Image Models Multi-Label Classification [software]. GitHub. https://github.com/yang-ruixin/PyTorch-Image-Models-Multi-Label-Classification. (Accessed 21 April 2026). Zhang, F., Zhou, Y., Mao, S., et al.,

    2026

  4. [2021]

    arXiv:2111.13606

    Conditional image generation with score-based diffusion models. arXiv:2111.13606. Black Forest Labs,

    arXiv

  5. [2024]

    (Accessed 21 April 2026)

    FLUX.1 [software].https://bfl.ai/. (Accessed 21 April 2026). Chen, H., Zhao, W., Zhang, R., Li, N., Li, D.,

    2026

  6. [2025]

    IEEE Geosci

    Multiple object tracking in video SAR: A benchmark and tracking baseline. IEEE Geosci. Remote Sens. Lett. 22, Art. no. 4010905. doi: 10.1109/LGRS.2025.3592711. Chen, J., Yu, J., Ge, C., et al.,

    2025

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.