Pith. sign in

REVIEW 4 major objections 7 minor 40 references

Quantitative Comparison of Fine-Tuning Techniques for Pretrained Latent Diffusion Models in the Generation of Unseen SAR Images

T0 review · 4 major / 7 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read A pretrained text-to-image diffusion model can be tuned to compose realistic, never-before-seen SAR radar scenes.

desk verdict Plausible hybrid fine-tuning recipe for SDXL-to-SAR, but the headline KL metric is partly trained into the best model and no uncertainty bounds are given. read the letter →

arxiv 2506.13307 v2 pith:5SS3EB23 submitted 2025-06-16 cs.CV cs.AI

classification cs.CVcs.AI
keywords syntheticapertureradarlatentdiffusionmodelStableXLLoRAtext-to-imagegenerationSARimagesynthesisparameter-efficientfine-tuningSAR-CLIPevaluation
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

This paper tries to establish that a large pretrained text-to-image diffusion model can be moved from ordinary photographs to Synthetic Aperture Radar (SAR) imagery without losing its ability to understand language and compose scenes. The authors adapt Stable Diffusion XL to 100,000 X-band radar patches at 40 cm resolution and systematically compare full fine-tuning against low-rank LoRA adapters in the UNet, text encoders, and VAE. They argue the winning recipe is a hybrid: fully fine-tune the UNet to learn radar-specific speckle, geometry, and amplitude statistics, apply LoRA to the text encoders to keep language knowledge intact, and learn a dedicated token embedding. If right, this would make radar imagery a controllable, prompt-driven generative medium, with practical value for data augmentation and rare-scene simulation in Earth observation.

What carries the argument

The load-bearing object is Stable Diffusion XL's latent diffusion pipeline: a Variational Autoencoder maps images into latents, two text encoders produce token-level and global embeddings, and a UNet denoises latents conditioned on those embeddings through cross-attention. The fine-tuning recipe modifies this pipeline in three modular ways: full gradient updates of the UNet weights, LoRA low-rank adapter matrices on the text-encoder attention projections, and a learnable embedding for a new <SAR> token, plus a short low-noise refinement of the VAE decoder with a KL distribution-matching loss. LoRA updates frozen weights additively with rank-constrained factors, keeping adapter capacity small, while full UNet tuning absorbs the statistical shift that SAR's heavy-tailed amplitude distributions and slant-range geometry impose.

What would settle it

Take the 330 evaluation prompts, replace the auto-generated optical captions with independently verified radar-aware captions, and recompute the SAR-CLIP rank and similarity scores for the configurations in Table 2; if the ordering no longer puts the hybrid configurations on top, the reported advantage is an artifact of caption noise. A second check is to verify by geolocation or patch retrieval that the claimed novel scenes, such as a circular structure in a city center or a river through dense forest, are genuinely absent from the training set.

Watch

Extended reading notes

Core claim

The central claim is that a vision-language foundation model pretrained on optical images can be adapted to generate physically grounded SAR scenes, and that the optimal configuration is neither all-full nor all-LoRA but a hybrid. Full fine-tuning of the UNet captures SAR's low-level statistics (Rayleigh-distributed amplitudes, speckle, side-looking geometry), while LoRA on both text encoders preserves spatial relations and object semantics, and a learned <SAR> token shifts the caption embedding toward radar style. The paper reports that the hybrid model heart-rose-2 reaches a KL divergence of 0.23 against real amplitude distributions and attains the best SAR-CLIP ranks in most evaluated categories, while composing scenes the training set did not contain, such as a circular structure in a city center or a river through dense forest.

Load-bearing premise

The results stand or fall on the assumption that the captions automatically generated from optical imagery are faithful descriptions of the co-aligned SAR patch; if those captions miss radar-specific content, the semantic alignment scores no longer measure how well the generated SAR image matches the prompt.

Editorial extensions

If this is right

  • Text prompts alone can control SAR scene composition, including spatial relations like 'near' and 'along', at 40 cm resolution.
  • The same adapted model can be reused as an image conditioner: it sharpens 1.35 m TerraSAR-X images to 40 cm and adds physically plausible texture to outputs of the EMPRISE physics-based simulator.
  • Rare or operationally relevant scenes absent from the training set, such as a bridge over a river or a circular structure in a city, can be synthesized on demand.
  • The adaptation runs on a single H100 GPU, so the procedure is reproducible for other resolutions or other latent diffusion backbones.

Reading between the lines

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

  • Inference: the same division of labor seen here—full backbone tuning for low-level statistics, low-rank adapters for language semantics, and a learned token for domain identity—likely transfers to other non-optical modalities with heavy-tailed statistics, such as sonar, lidar echo, or medical ultrasound.
  • Inference: because semantic evaluation depends on auto-generated optical captions, re-running the rankings with manually verified or radar-native captions is a natural stress test; if configuration ordering changes, part of the hybrid's reported advantage is an artifact of caption noise.
  • Inference: the learned <SAR> token probably behaves like textual inversion, meaning 'a <SAR> image of X' should inherit radar statistics even when X is an unseen concept; that compositional property is directly testable.
  • Inference: the claim that generated scenes are 'unseen' relies on absence from the training split, so checking by geolocation or patch retrieval whether the composed scenes actually appear in training data would sharpen or weaken the out-of-distribution claim.
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.

Referee Report

4 major / 7 minor

Summary. The paper adapts Stable Diffusion XL to high-resolution X-band SAR imagery. It curates 100k 1024x1024 patches from ONERA SETHI data, auto-generates captions from co-aligned optical images with CogVLM2, and compares fine-tuning configurations that vary full UNet fine-tuning, LoRA on the two text encoders, noise offset, batch size, VAE decoder refinement, and a learnable <SAR> token. Evaluation combines KL divergence between real and generated amplitude histograms, GLCM texture features, and a SAR-specialized CLIP rank/similarity. The central claim is that the hybrid configuration with full UNet tuning, LoRA on both text encoders, and the learned <SAR> token (heart-rose-2) best preserves SAR geometry and texture while maintaining prompt fidelity, and can compose scenes not present in the training set.

Significance. If the result holds, the paper would make a useful contribution: it demonstrates that a text-to-image foundation model can be transferred to a non-optical, physics-driven modality, and it provides a systematic comparison of full fine-tuning versus LoRA across model components. The proposed evaluation framework (KL, GLCM, SAR-CLIP) and the practical applications to TerraSAR-X and EMPRISE refinement are also valuable. However, the headline KL metric is also used as a training loss in the refinement phase, so the most emphasized advantage of the best model is not independent; the single-seed comparisons without uncertainty intervals further limit the strength of the ranking. The paper is within scope and the central idea is defensible, but the evidence as presented does not fully establish the ordering of configurations.

major comments (4)
  1. [5.5-5.6, Eqs. (10)-(11), Table 6] The primary statistical-fidelity metric is used as a training objective. Section 5.5 introduces L_refine = L_base + lambda_KL D_KL(Preal || Pgen), and Section 5.6 states that the <SAR> token embedding is updated with gradients from the diffusion objective and the distributional terms. The evaluation metric in Section 4.3, Eq. (4), is the same KL divergence between real and generated amplitude distributions. The best model heart-rose-2 is therefore trained to minimize the criterion on which it is judged, so its KL advantage over whale-north-8 (0.23 vs 0.35) is at least partially a consequence of optimizing the evaluation target rather than an independent measure of SAR realism. I recommend evaluating at least one configuration trained without the KL term, or reporting KL only as a secondary metric, so that the comparison is not self-referential.
  2. [5.2-5.6, Tables 2-6] All comparisons are single-seed, single-checkpoint runs at epoch 8, with no confidence intervals or significance tests. The differences between top models are small (e.g., CLIP rank 1.61 vs 1.74; KL 0.33 vs 0.35) and could plausibly lie within run-to-run variation. The paper selects the best configuration based on these point estimates, so the central claim is fragile without repeated seeds or bootstrap estimates. Please provide at least 3-5 seeds per configuration, or bootstrap over the 330 evaluation images, and state whether the reported differences are statistically distinguishable. The per-category results in Table 7 likewise lack any uncertainty quantification.
  3. [3 and 4.3] The prompt-fidelity evaluation depends on captions whose reliability is not established for the SAR modality. Section 3 states that textual descriptions were automatically generated from the co-aligned optical images using CogVLM2, and the same captions are used as prompts for generation and as ground truth for the CLIP rank and similarity scores in Section 4.3. If the optical captions miss radar-specific content, contain scene-label errors (Section 3 acknowledges possible mislabeling), or use generic phrasing that does not discriminate physical layout, the CLIP scores no longer measure SAR-prompt alignment. Please validate caption quality on a human-annotated SAR subset, or at least quantify caption noise and show that the ranking of configurations is robust to it.
  4. [4.3, Figure 15] The textural component of the central claim is not quantitatively established across configurations. The GLCM analysis is described in Section 4.3, but the results are presented only as qualitative panels in Figure 15 for heart-rose-2 versus real SAR images, and the caption mentions epoch 10 while the quantitative tables are at epoch 8. There is no per-configuration table of the four Haralick features. Please report the GLCM descriptors with means and standard deviations for all compared models (or at least for the configurations in Tables 2 and 6), and state whether the differences are statistically significant; otherwise the claim that the hybrid strategy best preserves SAR texture is not supported by the evidence as presented.
minor comments (7)
  1. [5.1 vs 5.2, Figure 5 and Table 2] There is a numeric inconsistency for the same configuration: Figure 5 reports rain-beach-6 with KL = 0.17, while Table 2 reports rain-beach-6 with KL = 1.17. Please reconcile these values or clarify that they refer to different checkpoints or evaluation conditions.
  2. [Tables 2 and 4] The model soleil-up-7 is reported with KL = 0.42 in Table 2 and KL = 0.43 in Table 4 under what appears to be the same configuration and epoch. Please specify the exact evaluation protocol or checkpoint used for each table so that the numbers are reproducible.
  3. [Figure 15] The caption states '(epoch 10)', while the rest of the experimental section reports results at epoch 8. Please clarify which checkpoint was used for the GLCM analysis and whether the visual comparison is representative of the quantitative comparisons.
  4. [3 and 5] Section 3 says captions were generated only for a subset of geo-aligned optical pairs, while Section 5 refers to a training dataset of 100,000 SAR image-caption pairs. Please clarify how many of the 100k samples have captions and how the remaining samples are used during training.
  5. [Appendix B] The caption of Figure B.2 says 'for 9 different models', but the figure shows only three models (whale-north-8, whale-north-8-refined, heart-rose-2). Please correct the caption.
  6. [5.3] The statement 'Contrary to what has been observed in the literature' would benefit from a supporting citation or a more precise description of the specific prior observations being compared.
  7. [General] The paper does not state whether code, trained checkpoints, or the evaluation dataset will be released. A data and code availability statement would strengthen the reproducibility claims made in Section 5.

Circularity Check

1 steps flagged · score 6.0 of 10

The KL-divergence metric is also the training objective of the refinement phase, so the reported fidelity gain of the best model is partly by construction.

  1. fitted input called prediction [Section 4.3 Eq. (4); Section 5.5 Eqs. (10)-(11); Tables 5 and 6]
    "In addition, we add a Kullback–Leibler (KL) divergence term to the loss to minimize the divergence between the amplitude distribution of the generated image ˆx and that of the target image x: LKL = DKL(Preal(x)∥ Pgen(ˆx)). (10) The total loss used during refinement is Lrefine = Lbase + λKLLKL, (11)"

    Equation (10) is the same D_KL(P∥Q) functional as the primary statistical-fidelity metric defined in Eq. (4). The refinement phase minimizes this quantity, and Table 5 then reports the refined model's lower KL (0.33 vs 0.35) as evidence of improved fidelity. The improvement is therefore partly a consequence of optimizing the evaluation target, not an independent measurement. The same applies to heart-rose-2 in Section 5.6 and Table 6, whose token embedding is updated with "our distributional terms" and then credited with the lowest KL (0.23). CLIP rank and GLCM metrics are not part of the training loss and provide independent partial support, so the circularity is partial rather than total.

full rationale

The paper's comparison is mostly self-contained: it compares fixed fine-tuning configurations with held-out generation and uses CLIP rank and GLCM features as external, non-optimized metrics. The one load-bearing circular step is the KL divergence. Section 5.5 adds D_KL to the refinement loss (Eqs. 10-11), and Sections 5.5-5.6 report reduced or lowest KL as evidence of realism (Tables 5-6). Since the evaluation KL of Eq. (4) is the same functional as the training term, the refined and <SAR>-token models are being graded on an objective they were explicitly trained to minimize. The reported improvements (0.35 to 0.33 and 0.35 to 0.23) are thus at least partly forced by construction. CLIP rank (Eqs. 5-6) and GLCM textural indicators are independent of the training losses and support the hybrid ranking, though GLCM results are not tabulated per configuration and no confidence intervals are given. Self-citations in Section 6 (Debuysère et al. 2024/2025, Trouve et al. 2024) concern applied ControlNet pipelines and are not load-bearing for the paper's central comparison. Overall: partial circularity in the headline statistical-fidelity comparison, not a fully circular derivation.

Assumptions & free parameters 7 free parameters · 6 assumptions · 1 invented entities

The main inputs are the transferability of an optical pretrained prior, the caption-image pairing derived from optical imagery, and the validity of the statistical and CLIP metrics. The most consequential input is the KL divergence term in the refinement loss, which also appears as an evaluation metric.

free parameters (7)
  • Noise offset gamma = 0.035
    Eq. 8 adds gamma * delta to the forward noise; gamma was selected by comparing one run with and one without the offset, then fixed for all later experiments.
  • LoRA rank and alpha for text encoders = r = 8, alpha = 4
    Table 3 compares ranks and selects the lowest rank as best; this is a configuration parameter chosen by the evaluation, not derived.
  • UNet and text encoder learning rates = 5e-5 (UNet), 4e-5 (text encoders)
    Section 5 says these values were chosen based on empirical stability on one H100, i.e. hand-tuned constants.
  • KL loss weight lambda_KL = not reported
    Eq. 11 defines L_refine with lambda_KL but the paper never gives its value, leaving a central training hyperparameter unspecified.
  • Batch size = 16 default, 64 best
    Section 5.4 varies batch size with gradient accumulation and picks a mid-sized value as optimal; another configuration choice that affects scores.
  • Low-noise refinement schedule fraction = last 15% of reverse diffusion, one epoch
    Section 5.5 restricts training timesteps to the final 15% of the reverse schedule; chosen by hand to avoid latent drift.
  • Evaluation saturation exclusion threshold = up to 3% of pixels excluded
    Section 4.3 removes saturated pixels before computing KL; the threshold affects all reported distribution distances.
assumptions (6)
  • domain assumption Stable Diffusion XL pretrained on optical text-image pairs retains semantic and compositional priors that transfer to SAR.
    The whole approach depends on this transfer (Introduction, Sections 4.2 and 7); if the prior does not transfer, the fine-tuned model is just a SAR autoencoder with no prompt control.
  • domain assumption Captions auto-generated by CogVLM2 from optical imagery accurately describe the geo-aligned SAR scene content.
    Section 3 states that 'textual descriptions were automatically generated for the optical images' and used as text labels for SAR patches. This is load-bearing for all prompt fidelity metrics.
  • domain assumption The KL divergence between real and generated amplitude histograms, after excluding saturated pixels and renormalizing, is a valid measure of radiometric realism.
    Section 4.3 defines the evaluation metric; Section 5.5 also uses the same divergence as a training loss, so both the metric and the objective share this assumption.
  • domain assumption The authors' fine-tuned SAR-CLIP model is a valid proxy for semantic alignment between prompts and generated SAR images.
    Section 4.3 uses ranking and cosine similarity from this model without human validation or comparison against a published SAR-text retrieval benchmark.
  • domain assumption The random split into train, validation, and test sets prevents geographic or scene-level leakage between training and test patches.
    If patches from the same large SETHI acquisition appear in both training and test sets, the reported realism and prompt fidelity would be inflated. The paper mentions excluding overlapping scenes but does not describe a flight or strip level split.
  • domain assumption SAR amplitude pixel values follow a Rayleigh distribution, used here as the physical target for statistical evaluation.
    Section 3 states the Rayleigh assumption; it is standard for single-look amplitude but the curated dataset at 40 cm may have texture correlation that alters the exact distribution.
invented entities (1)
  • The <SAR> token embedding
    purpose: A new vocabulary token with a learnable embedding that replaces 'satellite view' with 'A <SAR> image of' to bias the frozen text encoders toward radar statistics.
    It is learned jointly with the model during training and only evaluated on the authors' own metrics; there is no external benchmark showing that the token carries SAR meaning outside this pipeline.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantitative Comparison of Fine-Tuning Techniques for Pretrained Latent Diffusion Models in the Generation of Unseen SAR Images." pith.science (2026). https://pith.science/paper/5SS3EB23

@misc{pith2026250613307,
  author       = {Pith},
  title        = {Pith review of: Quantitative Comparison of Fine-Tuning Techniques for Pretrained Latent Diffusion Models in the Generation of Unseen SAR Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5SS3EB23}},
  note         = {Machine review of arXiv:2506.13307}
}
read the original abstract

We present a framework for adapting a large pretrained latent diffusion model to high-resolution Synthetic Aperture Radar (SAR) image generation. The approach enables controllable synthesis and the creation of rare or out-of-distribution scenes beyond the training set. Rather than training a task-specific small model from scratch, we adapt an open-source text-to-image foundation model to the SAR modality, using its semantic prior to align prompts with SAR imaging physics (side-looking geometry, slant-range projection, and coherent speckle with heavy-tailed statistics). Using a 100k-image SAR dataset, we compare full fine-tuning and parameter-efficient Low-Rank Adaptation (LoRA) across the UNet diffusion backbone, the Variational Autoencoder (VAE), and the text encoders. Evaluation combines (i) statistical distances to real SAR amplitude distributions, (ii) textural similarity via Gray-Level Co-occurrence Matrix (GLCM) descriptors, and (iii) semantic alignment using a SAR-specialized CLIP model. Our results show that a hybrid strategy-full UNet tuning with LoRA on the text encoders and a learned token embedding-best preserves SAR geometry and texture while maintaining prompt fidelity. The framework supports text-based control and multimodal conditioning (e.g., segmentation maps, TerraSAR-X, or optical guidance), opening new paths for large-scale SAR scene data augmentation and unseen scenario simulation in Earth observation.

Figures

Figures reproduced from arXiv: 2506.13307 by the authors.

Figure 1
Figure 1. Pairs of optical (ground plan) and SAR (slant-range plan) images. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. (a) Training dataset labels repartition (b) Dataset repartition; train, [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Forward and Reverse Process in Stable Di [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Random timesteps sampling over a batch during training epochs [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Comparison of trainings - at epoch 8 - with and without noise o [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Comparison of KL distances probability density distribution between [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Comparison of image generated (1024x1024px at 40cm) during [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Comparison of real and generated images (1024x1024px at 40cm) - at [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Mean cosine distance between image-text pairs, compared to real ones [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Comparison of KL distance distributions between 330 real and [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 11
Figure 11. Figure 11: Comparison of KL distance distributions between 330 real and [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 12
Figure 12. Figure 12: Comparison of generated images (1024x1024px at 40cm) — at epoch 8 [PITH_FULL_IMAGE:figures/full_fig_p011_12.png]
Figure 14
Figure 14. Figure 14: Comparison of generated images (1024x1024px at 40cm) — at [PITH_FULL_IMAGE:figures/full_fig_p012_14.png]
Figure 15
Figure 15. Figure 15: GLCM texture metrics (epoch 10) for real vs generated SAR images [PITH_FULL_IMAGE:figures/full_fig_p013_15.png]
Figure 16
Figure 16. Figure 16: Generated images (1024x1024px at 40cm) from the [PITH_FULL_IMAGE:figures/full_fig_p014_16.png]
Figure 18
Figure 18. Figure 18: Enhancing ONERA’s physics-based simulator EMPRISE Images [PITH_FULL_IMAGE:figures/full_fig_p015_18.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 12 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in ":" * " " * FUNCTION f...

  2. [2]

    Agrawal and R

    K. Agrawal and R. Banerjee. Synthetic art generation and deepfake detection: A study on jamini roy inspired dataset. TechRxiv, Mar. 2025. doi:10.36227/techrxiv.174119231.19482547/v1. URL http://dx.doi.org/10.36227/techrxiv.174119231.19482547/v1

  3. [3]

    S. Auer, R. Bamler, and P. Reinartz. Raysar - 3d sar simulator: Now open source. In 2016 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), pages 6730--6733, 2016. doi:10.1109/IGARSS.2016.7730757

  4. [4]

    Baqu \'e , P

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

  5. [5]

    COCHIN, P

    C. COCHIN, P. POULIGUEN, B. DELAHAYE, D. l. HELLARD, P. GOSSELIN, and F. AUBINEAU. Mocem - an 'all in one' tool to simulate sar image. In 7th European Conference on Synthetic Aperture Radar, pages 1--4, 2008

  6. [6]

    W. Dai, L. Lu, and Z. Li. Diffusion-based synthetic data generation for visible-infrared person re-identification, 2025. URL https://arxiv.org/abs/2503.12472

  7. [7]

    Debuysère, N

    S. Debuysère, N. Trouvé, N. Letheule, E. Colin, and O. Lévêque. Synthesizing sar images with generative ai: Expanding to large-scale imagery, October 2024. https://hal.science/hal-04786104

  8. [8]

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

    S. Debuysère, N. Trouvé, N. Letheule, O. Lévêque, and E. Colin. From spaceborn to airborn: Sar image synthesis using foundation models for multi-scale adaptation, 2025. URL https://arxiv.org/abs/2505.03844

Show all 40 references
  1. [9]

    Dettmers, A

    T. Dettmers, A. Pagnoni, A. Holtzman, and L. Zettlemoyer. Qlora: Efficient finetuning of quantized llms, 2023. URL https://arxiv.org/abs/2305.14314

  2. [10]

    R. Gal, Y. Alaluf, Y. Atzmon, O. Patashnik, A. H. Bermano, G. Chechik, and D. Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion, 2022. URL https://arxiv.org/abs/2208.01618

  3. [11]

    D. Gao, X. Wu, Z. Wen, Y. Xu, and Z. Chen. Few-shot sar vehicle target augmentation based on generative adversarial networks. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, X-1-2024: 0 83--90, 2024. doi:10.5194/isprs-annals-X-1-2024-83-202...

  4. [12]

    R. M. Haralick, K. Shanmugam, and I. Dinstein. Textural features for image classification. IEEE Transactions on Systems, Man, and Cybernetics, SMC-3 0 (6): 0 610--621, 1973. doi:10.1109/TSMC.1973.4309314

  5. [13]

    W. Hong, W. Wang, M. Ding, W. Yu, Q. Lv, Y. Wang, Y. Cheng, S. Huang, J. Ji, Z. Xue, L. Zhao, Z. Yang, X. Gu, X. Zhang, G. Feng, D. Yin, Z. Wang, J. Qi, X. Song, P. Zhang, D. Liu, B. Xu, J. Li, Y. Dong, and J. Tang. Cogvlm2: Visual language models for image and video understan...

  6. [14]

    E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen. Lora: Low-rank adaptation of large language models, 2021. URL https://arxiv.org/abs/2106.09685

  7. [15]

    M. Jia, L. Tang, B.-C. Chen, C. Cardie, S. Belongie, B. Hariharan, and S.-N. Lim. Visual prompt tuning, 2022. URL https://arxiv.org/abs/2203.12119

  8. [16]

    Khanna, P

    S. Khanna, P. Liu, L. Zhou, C. Meng, R. Rombach, M. Burke, D. Lobell, and S. Ermon. Diffusionsat: A generative foundation model for satellite imagery, 2024. URL https://arxiv.org/abs/2312.03606

  9. [17]

    Kirillov, E

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y. Lo, P. Dollár, and R. Girshick. Segment anything, 2023. URL https://arxiv.org/abs/2304.02643

  10. [18]

    B. F. Labs. Flux. https://github.com/black-forest-labs/flux, 2024

  11. [19]

    C. Liu, K. Chen, R. Zhao, Z. Zou, and Z. Shi. Text2earth: Unlocking text-driven remote sensing image generation with a global-scale dataset and a foundation model, 2025. URL https://arxiv.org/abs/2501.00895

  12. [20]

    W. Liu , Y. Zhao , M. Liu , L. Dong , X. Liu , and M. Hui . Generating simulated SAR images using Generative Adversarial Network . In A. G. Tescher , editor, Applications of Digital Image Processing XLI, volume 10752 of Society of Photo-Optical Instrumentation Engineers (SPIE)...

  13. [21]

    Z. Ma, X. Xiao, S. Dong, P. Wang, H. Wang, and Q. Pan. Sarchat-bench-2m: A multi-task vision-language benchmark for sar image interpretation, 2025. URL https://arxiv.org/abs/2502.08168

  14. [22]

    L. Pang, X. Cao, D. Tang, S. Xu, X. Bai, F. Zhou, and D. Meng. Hsigene: A foundation model for hyperspectral image generation, 2024. URL https://arxiv.org/abs/2409.12470

  15. [23]

    Podell, Z

    D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. Müller, J. Penna, and R. Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. URL https://arxiv.org/abs/2307.01952

  16. [24]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models, 2022. URL https://arxiv.org/abs/2112.10752

  17. [25]

    N. Ruiz, Y. Li, V. Jampani, Y. Pritch, M. Rubinstein, and K. Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation, 2023. URL https://arxiv.org/abs/2208.12242

  18. [26]

    Saharia, W

    C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. Denton, S. K. S. Ghasemipour, B. K. Ayan, S. S. Mahdavi, R. G. Lopes, T. Salimans, J. Ho, D. J. Fleet, and M. Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. URL https://arxiv.or...

  19. [27]

    D. Tang, X. Cao, X. Hou, Z. Jiang, J. Liu, and D. Meng. Crs-diff: Controllable remote sensing image generation with diffusion model, 2024. URL https://arxiv.org/abs/2403.11614

  20. [28]

    C. Team. Chameleon: Mixed-modal early-fusion foundation models, 2025. URL https://arxiv.org/abs/2405.09818

  21. [29]

    Trouve, N

    N. Trouve, N. Letheule, O. Leveque, I. Rami, and E. 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

  22. [30]

    Q. Wang, Y. Fan, J. Bao, H. Jiang, and Y. Song. Bora: Bi-dimensional weight-decomposed low-rank adaptation, 2024. URL https://arxiv.org/abs/2412.06441

  23. [31]

    Woollard, D

    M. Woollard, D. Blacknell, H. Griffiths, and M. A. Ritchie. Sarcastic v2.0—high-performance sar simulation for next-generation atr systems. Remote Sensing, 14 0 (11), 2022. ISSN 2072-4292. doi:10.3390/rs14112561. URL https://www.mdpi.com/2072-4292/14/11/2561

  24. [32]

    H. Ye, J. Zhang, S. Liu, X. Han, and W. Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models, 2023. URL https://arxiv.org/abs/2308.06721

  25. [33]

    J. Yu, Z. Wang, V. Vasudevan, L. Yeung, M. Seyedhosseini, and Y. Wu. Coca: Contrastive captioners are image-text foundation models, 2022. URL https://arxiv.org/abs/2205.01917

  26. [34]

    L. Yu, B. Shi, R. Pasunuru, B. Muller, O. Golovneva, T. Wang, A. Babu, B. Tang, B. Karrer, S. Sheynin, C. Ross, A. Polyak, R. Howes, V. Sharma, P. Xu, H. Tamoyan, O. Ashual, U. Singer, S.-W. Li, S. Zhang, R. James, G. Ghosh, Y. Taigman, M. Fazel-Zarandi, A. Celikyilmaz, L. Zet...

  27. [35]

    Z. Yu, C. Liu, L. Liu, Z. Shi, and Z. Zou. Metaearth: A generative foundation model for global-scale remote sensing image generation, 2024. URL https://arxiv.org/abs/2405.13570

  28. [36]

    Zhang, A

    L. Zhang, A. Rao, and M. Agrawala. Adding conditional control to text-to-image diffusion models, 2023. URL https://arxiv.org/abs/2302.05543

  29. [37]

    Zhang, T

    Z. Zhang, T. Zhao, Y. Guo, and J. Yin. Rs5m and georsclip: A large-scale vision- language dataset and a large vision-language model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing, 62: 0 1–23, 2024. ISSN 1558-0644. doi:10.1109/tgrs.2024.3449154. URL http...

  30. [38]

    K. Zhou, J. Yang, C. C. Loy, and Z. Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130 0 (9): 0 2337–2348, July 2022. ISSN 1573-1405. doi:10.1007/s11263-022-01653-1. URL http://dx.doi.org/10.1007/s11263-022-01653-1

  31. [39]

    L. Zou, H. Zhang, C. Wang, F. Wu, and F. Gu. Mw-acgan: Generating multiscale high-resolution sar images for ship detection. Sensors, 20 0 (22), 2020. ISSN 1424-8220. doi:10.3390/s20226673. URL https://www.mdpi.com/1424-8220/20/22/6673

  32. [40]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.