Pith. sign in

REVIEW 4 major objections 8 minor 50 references

From one text prompt, a single model can generate optical, infrared, and SAR remote sensing images that share the same scene content and spatial layout.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A LoRA-based contrastive parameter disentanglement method plus query-key structure transfer generates semantically consistent, structurally aligned OPT/IR/SAR images from a single text prompt.

T0 review reviewed 2026-07-30 challenge →

load-bearing objection Clever LoRA-core disentanglement for single-prompt OPT/IR/SAR, but a third of the WHU numbers rest on IR data the cited dataset does not contain. the 4 major comments →

arxiv 2607.23673 v1 pith:ZKPWOOVB submitted 2026-07-26 cs.CV

Contrastive Parameter Disentanglement for Multi-modal Remote Sensing Image Generation

Yu Zhang , Wenda Zhao , Haojun Tang , Haipeng Wang This is my paper
classification cs.CV
keywords multi-modal remote sensingtext-to-image generationcontrastive parameter disentanglementLoRAquery-key structure transferoptical infrared SARdiffusion models
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Current remote sensing generators only make one kind of image at a time, so they never use the fact that optical, infrared, and radar views of the same place carry complementary information. This paper claims that the low-rank adapters inside a diffusion model already split naturally into a shared semantic part and modality-specific parts, and that locking the shared part inside a carefully chosen orthogonal core of the network weights lets one text prompt drive all three modalities at once. A second inference-time step copies structural attention from an optical anchor into the other modalities so the layouts stay aligned. If the claim holds, practitioners can synthesize matched multimodal training sets from captions alone, improving both generation metrics and real-image classification when only synthetic data are used for training. The practical payoff is denser, weather-robust earth-observation data without collecting every sensor for every scene.

Core claim

The authors establish that LoRA adapters inside a stable-diffusion UNet already exhibit a functional split—matrix A is cross-modally similar while matrices B diverge—and that restricting adaptation to the effective-rank core subspace obtained from singular-value entropy lets A be contrastively trained for invariant semantics and separate B branches be trained for optical, infrared, and SAR attributes. Coupled with query-key structure transfer at inference, the resulting system produces semantically consistent and spatially aligned multimodal images from a single text prompt, outperforming single-modality generators and yielding higher downstream classification accuracy on real test images.

What carries the argument

Contrastive parameter disentanglement (CPD): LoRA factors are placed inside the orthogonal core subspace of each UNet attention weight; A is optimized with a multimodal contrastive loss to extract shared semantics, while multiple B matrices reconstruct modality attributes; query-key structure transfer then copies the optical self-attention map to the other modalities during sampling.

Load-bearing premise

The pattern that LoRA matrix A stays similar across modalities while B diverges, together with the core subspace defined by singular-value entropy, is a stable and general basis for cleanly separating shared meaning from sensor-specific appearance.

What would settle it

Retrain the same pipeline on weakly paired or unpaired multimodal collections, or swap the backbone for a non-SD UNet; if cross-modal cosine similarity of A collapses or FID/CS gains disappear, the central separation claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A single caption can simultaneously produce matched optical, infrared, and SAR training images without separate generators.
  • Synthetic multimodal sets generated this way raise real-image classification accuracy when classifiers are trained only on the synthetic data.
  • Structural alignment across modalities becomes available at inference without extra image-level supervision or paired layout maps.
  • Earth-observation pipelines that need all-weather coverage can expand scarce IR/SAR collections from abundant optical captions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the A/B split is truly backbone-agnostic, the same contrastive core-subspace recipe could be ported to video or hyperspectral diffusion models that also need shared content plus sensor-specific rendering.
  • Failure modes on unpaired data would directly test whether the method needs true multimodal correspondence or only caption-level class labels.
  • The optical-as-anchor choice in structure transfer suggests a natural extension: choosing the most reliable modality per scene (e.g., SAR under cloud) as the structural prior.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 8 minor

Summary. The paper proposes a text-to-multimodal remote sensing image generation framework built on SDv1.5. Its core observation is that in independently trained LoRA adapters, the A matrices exhibit high cross-modal cosine similarity while B matrices diverge (Fig. 2a), motivating a Contrastive Parameter Disentanglement (CPD) module: each UNet attention weight is SVD-decomposed, an effective-rank core subspace is extracted via singular-value entropy (Eqs. 5–8), and the LoRA adapter is constrained by the resulting orthogonal bases. A two-stage Disentangled Optimization Strategy first trains a shared A with a multimodal supervised-contrastive loss L_sd (Eqs. 14–16) to capture modality-invariant semantics, then trains per-modality B_i with a diffusion reconstruction loss L_ar under composite text+attribute prompts (Eqs. 17–18). At inference, a Query-Key Structure Transfer (QKST) mechanism copies the OPT branch's self-attention maps into the IR and SAR branches at every denoising step (Eqs. 19–23) to enforce structural alignment. Experiments on WHU-OPT-SAR and a newly assembled OSI dataset report state-of-the-art IS/FID/CS across all three modalities (Table 2), module and variant ablations (Tables 3–6), and a train-on-synthetic/test-on-real downstream classification study (Table 7).

Significance. If the claims hold, the paper makes three contributions of genuine interest: (i) it formulates text-to-multimodal remote sensing generation (OPT/IR/SAR from one prompt), a task that prior single-modality RS generators do not address; (ii) the CPD construction — anchoring LoRA adaptation in an entropy-derived effective-rank core subspace with a two-stage contrastive/reconstruction schedule — is a concrete, reproducible mechanism with thorough ablations (Tables 3–6) including the w/o CSE, w/o OB, w/o DOS variants and a t-SNE disentanglement visualization; (iii) the train-on-synthetic/test-on-real classification study (Table 7) is a meaningful, falsifiable downstream check beyond FID/IS. The QKST inference-time mechanism is simple and the SA-vs-CA transfer ablation (Table 5, Fig. 12) is a nice diagnostic. However, the significance of the three-modality headline is currently capped by unresolved data-provenance and evaluation-validity questions detailed below.

major comments (4)
  1. [§4.1, Tables 2–3] §4.1 states that WHU-OPT-SAR [38] is used for training and that 'each modality contains 7,000 images,' and Tables 2–3 report IS/FID/CS for the IR modality on this dataset (e.g., IR FID 27.31). However, reference [38] (MCANet) is a paired OPTICAL+SAR segmentation dataset; it contains no infrared channel, and no IR source or synthesis step is described anywhere in the manuscript. Computing FID requires a real IR reference set, so every IR number on WHU-OPT-SAR is currently unverifiable: either an unstated IR dataset/synthesis procedure exists, or the column is mislabeled. Relatedly, [38] is a segmentation dataset without text captions, yet the whole pipeline is text-conditioned and the CS metric requires prompts — the caption provenance (template? captioning model? manual?) is never given. Since the IR modality is one third of the central claim ('simultaneous OPT/IR/SAR generation'), theIR
  2. [§3.4 (Eq. 22), §4.1 (Table 1), §4.2] Nowhere in the pipeline do pixel-paired OPT/IR/SAR triplets exist: WHU-OPT-SAR pairs only OPT and SAR (and per Major 1, its IR is unexplained), and OSI is explicitly assembled from unpaired heterogeneous sources (NWPU-RESISC45 for OPT; SARDet/BRIGHT/FUSAR-Ship for SAR; DroneVehicle/VEDAI for IR — different sensors, regions, resolutions, and class distributions per Table 1). Two consequences. (a) 'Structurally aligned' is enforced by construction: Eq. (22) hard-copies the OPT self-attention map into the IR/SAR branches, and alignment is never measured against ground truth — the only quantitative proxy is the CLIP score, whose image encoder was trained on natural RGB and is of doubtful validity for SAR/IR (CS deltas of ~0.01–0.02 carry the claim in Table 2). A quantitative alignment evaluation is needed on data where ground truth exists — e.g., the paired OPT/SAR portion of WHU: cross-mod
  3. [§3.2 (Property 2, Eq. 13), Fig. 2(a)] The 'theoretical analysis' underpinning the CPD design is loose at the exact point where it is load-bearing. In Property 2(i), span(P_in) and span(P_out) are subspaces of R^n (row space) and R^m (column space) respectively — objects in different spaces — so the claim that their inequality 'naturally ensures that semantic and attribute updating operate on distinct orthogonal spaces' does not follow as stated. In Eq. (13), eliminating the U-terms via orthonormality is correct algebra, but the conclusion that 'independently optimizing the B_i suppresses the correlation (B_i)^T B_j' is asserted without argument; nothing in the L_ar objective (Eq. 18) penalizes cross-branch correlation. Likewise, the motivating observation in Fig. 2(a) — A matrices similar across modalities, B matrices divergent — lacks a control: independently trained LoRAs with A initialized from the same distribution (or a
  4. [§4.4 (Table 7), Table 3 vs. Table 5] The text states 'our method outperforms all competing methods across all modalities and synthesis scales,' but its own table contradicts this: on OPT at 100% scale with VGG19, OTD-GAN achieves 0.8250 vs. Ours 0.8036. The claim should be corrected to the average-case statement the data supports. Two further numerical inconsistencies need reconciliation: (a) Table 5's SAR row for 'w SA (Ours)' reports CS 0.2766, while the identical full-model configuration in Table 3 reports 0.2855; (b) in Table 5, 'w CA' attains higher CS than 'w SA (Ours)' on both IR (0.2857 vs 0.2866 is fine) and SAR (0.2825 vs 0.2766), yet the paper bolds its own variant as best — on SAR CS the bolded value is not the best in its column. These are small deltas, but the paper's alignment argument leans on CS, so the tables must be internally consistent.
minor comments (8)
  1. [Eq. (23)] The DDIM-style update appears to use per-step α_t where the cumulative ᾱ_t of Eqs. (2)–(3) is intended, and the stochastic term σ_t ξ present in Eq. (3) is dropped; please check for a transcription error.
  2. [Eq. (15)] sim(·,·,c) is described as 'cosine similarity conditioned on the text prompt c,' but how c enters the cosine computation is never defined. Please make this explicit. The supervised contrastive form of Eq. (15) should also cite Khosla et al. (SupCon).
  3. [§4.2–4.3] All quantitative results are single-run; no error bars or multi-seed variance are reported. Several OSI deltas (e.g., OPT IS 5.642 vs 5.622 for CRS-Diff) are within plausible seed noise. At minimum, report variance for the main Table 2 result.
  4. [§4.1 (metrics)] CLIP score is used to claim semantic alignment for SAR/IR, but CLIP's encoder is RGB-natural-image trained. Consider additionally reporting a remote-sensing text-image model (e.g., RemoteCLIP) or a human/automated attribute check to support the CS conclusions.
  5. [§4.2.1] The sentence 'For IR and SAR modalities, previous methods produce images with compromised fine-grained details and texture fidelity...' is repeated twice in succession.
  6. [Table 2 (formatting)] OSI SAR entry for DiffusionSat runs together ('2.950 97.81'); also the GitHub URL in the abstract contains a stray space ('DUT -ZYu').
  7. [Fig. 2(a) (clarification)] Please state whether the LoRA A matrices in the Fig. 2(a) analysis were trained or held at initialization, the LoRA rank used, and how many layers/seeds the similarity pattern aggregates over.
  8. [§1, contribution bullets] The claim 'guaranteeing spatial structure consistency' overstates an inference-time heuristic; the ablation in Table 5 itself shows metric movement of mixed sign. Recommend softening to 'enforcing' or 'promoting' and reporting one instance-level alignment number.

Circularity Check

1 steps flagged

Empirical methods paper: generation metrics are external objectives; only the structural-alignment subclaim is partly true by the QKST construction.

specific steps
  1. self definitional [§3.4 Eqs. (21)–(22); also Abstract / §1 claim of structural alignment]
    "Rather than computing independent spatial attention per modality, QKST explicitly transfers the attention map A_opt_t,l to the IR and SAR branches, guiding their structural alignment to OPT branch. The output feature ˆf^m_t,l of the subsequent attention block is thus given by: ˆf^m_t,l = A_opt_t,l ⊗ V^m_t,l, m∈{opt,ir,sar}. ... ensuring structurally aligned multimodal generation"

    Cross-modal structural alignment is defined by copying the OPT QK attention map into the other modalities at every diffusion step. The paper then cites the resulting images as evidence of structural alignment. For that subclaim alone, the mechanism and the reported property coincide by construction; no independent structural metric against paired GT is given. This does not force FID/IS/CS or classification numbers.

full rationale

This is a standard empirical ML paper. The load-bearing results are IS/FID/CS on held-out prompts and downstream real-image classification accuracy when classifiers are trained only on synthetic images (Tables 2, 3, 7). Those metrics are external distributional/semantic objectives; nothing in L_sd, L_ar, the SVD core-subspace construction (Eqs. 4–9), or the two-phase DOS forces the reported scores by identity. The sole mild circularity is local to the structural-alignment claim: QKST hard-transfers the OPT self-attention map into the IR and SAR branches (Eqs. 21–22), so spatial correspondence is imposed by the inference rule rather than independently verified against paired ground truth. That is an inductive-bias design choice, not a fitted constant relabeled as a parameter-free prediction, and it does not underwrite the FID/IS/CS or classification gains. No self-citation uniqueness theorem, no ansatz smuggled from overlapping-author theory, and no renaming of a known closed-form result. Data-provenance concerns (IR channel on WHU-OPT-SAR) are correctness/validity issues, not circularity. Score 1 reflects one minor self-definitional step on a secondary claim.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 3 invented entities

The method rests on pretrained SD-v1.5, the LoRA factorization, an empirical A-vs-B similarity observation, entropy-based core rank, contrastive pairing of multimodal images that share class/text semantics, and the choice of optical self-attention as structural ground truth at inference. Free knobs (rank, stages, τ, guidance) and these domain assumptions carry the central claim more than any closed-form derivation.

free parameters (5)
  • LoRA rank r1 = 32
    Chosen by ablation among {16,32,64}; best reported at 32 and used for main results (Table 6).
  • contrastive temperature τ = 0.07
    Fixed default in multimodal contrastive loss L_c without extensive justification.
  • effective rank r = ceil(exp(H(W_l))) = layer-adaptive
    Per-layer core dimension from singular-value entropy; defines the orthogonal bases that constrain A/B (Eqs. 5–8).
  • two-stage training schedule and learning rates = 1e-5 / 1e-4; 10/20/100 epochs
    Domain fine-tune 10 epochs lr 1e-5; CPD phase I 20 epochs and phase II 100 epochs at lr 1e-4, batch 8—hand-chosen training recipe.
  • DDIM steps and CFG scale = 50 steps; CFG 7.5
    Inference hyperparameters fixed for all reported generations.
axioms (6)
  • ad hoc to paper LoRA ΔW=BA exhibits a functional dichotomy: A is a cross-modal invariant semantic anchor and B encodes modality-specific attributes.
    Motivated by cosine-similarity measurements on independently trained adapters (Fig. 2a, §1); treated as design principle for CPD rather than a proven general law.
  • domain assumption Core generative knowledge of each UNet attention weight lies in a low effective-rank subspace measurable by singular-value entropy; residual directions are task-irrelevant redundancy.
    Used to justify optimizing only inside W_core and orthogonal bases P_in/P_out (Fig. 2b, Eqs. 4–12).
  • domain assumption Self-attention QK maps dominate spatial structure of the diffusion trajectory; cross-attention mainly encodes coarse text-object binding.
    Justifies transferring only SA maps in QKST (Fig. 5, §3.4); ablation prefers SA-only over CA or CA+SA.
  • domain assumption Optical imagery is a reliable geometric anchor whose structural priors should be imposed on IR and SAR sampling trajectories.
    Hard-coded in QKST (Eqs. 21–22, Fig. 6); if OPT structure is wrong for radar/thermal geometry, alignment is forced incorrectly.
  • domain assumption Multimodal images sharing class/text labels provide positive pairs for invariant semantics under a single prompt.
    Required for L_sd construction on WHU-OPT-SAR and OSI (§3.3, §4.1).
  • domain assumption Standard latent diffusion reverse process and frozen VAE decoder of SD-v1.5 are adequate for RS OPT/IR/SAR synthesis after light domain fine-tuning.
    Backbone and training setup in §3.1 and §4.1.
invented entities (3)
  • Contrastive Parameter Disentanglement (CPD) module no independent evidence
    purpose: Parameter-level split of shared semantics (A) vs modality attributes (B_i) inside orthogonal core subspaces of UNet attention layers.
    Central architectural invention; evidence is internal ablations and metrics, not an external physical handle.
  • Query-Key Structure Transfer (QKST) no independent evidence
    purpose: Copy OPT self-attention maps to IR/SAR branches during sampling for cross-modal spatial alignment.
    Inference mechanism introduced in §3.4; validated only within this paper’s generations.
  • Disentangled Optimization Strategy (DOS) no independent evidence
    purpose: Two-phase training: contrastive freeze-B train-A, then freeze-A train-B_i with attribute-augmented prompts.
    Procedural invention tied to CPD; no independent external test outside reported experiments.

reviewed 2026-07-30 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Contrastive Parameter Disentanglement for Multi-modal Remote Sensing Image Generation." pith.science (2026). https://pith.science/paper/ZKPWOOVB

@misc{pith2026260723673,
  author       = {Pith},
  title        = {Pith review of: Contrastive Parameter Disentanglement for Multi-modal Remote Sensing Image Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZKPWOOVB}},
  note         = {Machine review of arXiv:2607.23673}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Existing remote sensing image generation methods are largely confined to single-modality synthesis and therefore fail to exploit the complementary information inherent in multimodal imagery. To address this limitation, we propose a contrastive parameter disentanglement framework for multimodal remote sensing image generation, which generates semantically consistent and structurally aligned images across multiple modalities, including optical, infrared, and synthetic aperture radar (SAR), from a single text prompt. Specifically, we introduce a contrastive parameter disentanglement module that disentangles shared semantics from modality-specific attributes at the parameter level within an orthogonal core subspace. Based on this module, we develop a disentangled optimization strategy that first constrains the parameter matrix A of the LoRA adapter to capture modality-invariant semantics through a multimodal contrastive objective and then guides multiple parameter matrices B to learn modality-specific attributes under text conditioning. This strategy enables the simultaneous generation of multimodal images with consistent semantic content and distinct modality characteristics. Furthermore, to ensure structural alignment across the generated images, we devise a query-key structure transfer mechanism that jointly models multimodal sampling trajectories during inference by transferring structural correlation priors from an anchor modality to the remaining modalities. Extensive experiments demonstrate that our method outperforms state-of-the-art remote sensing image generation approaches in terms of generation quality, semantic consistency, and structural alignment, while also achieving superior performance in the downstream object classification task.

Figures

Figures reproduced from arXiv: 2607.23673 by Haipeng Wang, Haojun Tang, Wenda Zhao, Yu Zhang.

Figure 1
Figure 1. Figure 1: Comparison of the generated results. The multimodal images generated by Text2Earth [ [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Statement of the motivation. (a) Similarity analysis of the parameter matrices [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The training framework of our proposed CPD module. The pre-trained parameters are first decomposed into a core parameter subspace to [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of the attribute reconstruction loss. We concatenate [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visualizations of QK attention activation maps across distinct UNet stages, including SA and CA heatmaps from the encoder (Down), middle (Mid), and decoder (Up) blocks. “The blue lake lies in an area of farmland” z UNet A t t e n tio n A t t e n tio n A t t e n tio n A t t e n tio n A t t e n tio n A t t e n tio n C P D C P D C P D C P D C P D C P D Query-Key Structure Transfer (QKST) [PITH_FULL_IMAGE:fig… view at source ↗
Figure 6
Figure 6. Figure 6: Overview framework of the multimodal generation inference. Leveraging only a single text prompt [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative generation results on the WHU-OPT-SAR dataset. From top to bottom are the OPT, IR, and SAR modalities, respectively. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative generation results on the OSI dataset. The generated images from top to bottom are the OPT, IR, and SAR modalities, respectively. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison generated results using LoRA [ [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Comparison results of different variants of CPD module. [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: t-SNE visualization of the feature disentanglement. (a) The conventional LoRA optimization paradigm [ [PITH_FULL_IMAGE:figures/full_fig_p012_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Comparison results of different variants of QKST. [PITH_FULL_IMAGE:figures/full_fig_p012_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references · 7 linked inside Pith

  1. [1]

    Ringmoe: Mixture-of-modality- experts multi-modal foundation models for universal remote sens- ing image interpretation,

    H. Bi, Y. Feng, B. Tonget al., “Ringmoe: Mixture-of-modality- experts multi-modal foundation models for universal remote sens- ing image interpretation,”IEEE T ransactions on Pattern Analysis and Machine Intelligence, pp. 1–18, 2025

    2025

  2. [2]

    Spectralgpt: Spectral remote sensing foundation model,

    D. Hong, B. Zhang, X. Liet al., “Spectralgpt: Spectral remote sensing foundation model,”IEEE transactions on pattern analysis and machine intelligence, vol. 46, no. 8, pp. 5227–5244, 2024

    2024

  3. [3]

    Focus-n-fix: Region-aware fine- tuning for text-to-image generation,

    X. Xing, A. Saha, J. Heet al., “Focus-n-fix: Region-aware fine- tuning for text-to-image generation,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2025, pp. 18 486–18 496

    2025

  4. [4]

    Conceptguard: Continual personalized text- to-image generation with forgetting and confusion mitigation,

    Z. Guo and T. Jin, “Conceptguard: Continual personalized text- to-image generation with forgetting and confusion mitigation,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 2945–2954

    2025

  5. [5]

    Text-guided diverse image synthesis for long-tailed remote sensing object classification,

    H. Tang, W. Zhao, G. Huet al., “Text-guided diverse image synthesis for long-tailed remote sensing object classification,”IEEE T ransactions on Geoscience and Remote Sensing, vol. 62, pp. 1–13, 2024

    2024

  6. [6]

    Text2earth: Unlocking text-driven remote sensing image generation with a global-scale dataset and a foundation model,

    C. Liu, K. Chen, R. Zhaoet al., “Text2earth: Unlocking text-driven remote sensing image generation with a global-scale dataset and a foundation model,”IEEE Geoscience and Remote Sensing Magazine, vol. 13, no. 3, pp. 238–259, 2025

    2025

  7. [7]

    Metaearth: A generative foundation model for global-scale remote sensing,

    Z. Yu, C. Liu, L. Liuet al., “Metaearth: A generative foundation model for global-scale remote sensing,”IEEE T rans. Pattern Anal. Mach. Intell., vol. 47, no. 3, pp. 1764–1781, 2025

    2025

  8. [8]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenzet al., “High-resolution image synthesis with latent diffusion models,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695

    2022

  9. [9]

    Remote sensing image gen- eration via object text decoupling,

    W. Zhao, Z. Zhang, F. Zhaoet al., “Remote sensing image gen- eration via object text decoupling,”IEEE T ransactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 12, pp. 11 834–11 847, 2025

    2025

  10. [10]

    Cc-diff++: Spatially controllable text-to-image synthesis for remote sensing with en- hanced contextual coherence,

    M. Zhang, Y. Liu, Y. Liu, Y. Zhao, and Q. Ye, “Cc-diff++: Spatially controllable text-to-image synthesis for remote sensing with en- hanced contextual coherence,”IEEE T ransactions on Geoscience and Remote Sensing, vol. 63, pp. 1–16, 2025

    2025

  11. [11]

    Mmo-ig: Multi-class and multi-scale object image generation for remote sensing,

    C. Yang, B. Zhao, Q. Zhou, and Q. Wang, “Mmo-ig: Multi-class and multi-scale object image generation for remote sensing,”IEEE T ransactions on Geoscience and Remote Sensing, 2025

    2025

  12. [12]

    Rsdiff: Remote sensing image gen- eration from text using diffusion model,

    A. Sebaq and M. ElHelw, “Rsdiff: Remote sensing image gen- eration from text using diffusion model,”Neural Computing and Applications, vol. 36, no. 36, pp. 23 103–23 111, 2024

    2024

  13. [13]

    Diffusionsat: A generative foundation model for satellite imagery,

    S. Khanna, P . Liu, L. Zhouet al., “Diffusionsat: A generative foundation model for satellite imagery,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  14. [14]

    Geosynth: Contextually-aware high-resolution satellite image synthesis,

    S. Sastry, S. Khanal, A. Dhakal, and N. Jacobs, “Geosynth: Contextually-aware high-resolution satellite image synthesis,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 460–470

    2024

  15. [15]

    Learning disentangled representations for perceptual point cloud quality assessment via mutual information minimization,

    Z. Shan, Y. Zhang, Y. Liu, and Y. Xu, “Learning disentangled representations for perceptual point cloud quality assessment via mutual information minimization,”Advances in Neural Information Processing Systems, vol. 37, pp. 84 060–84 080, 2024

    2024

  16. [16]

    Learning disen- tangled representation for one-shot progressive face swapping,

    Q. Li, W. Wang, C. Xu, Z. Sun, and M.-H. Yang, “Learning disen- tangled representation for one-shot progressive face swapping,” IEEE transactions on pattern analysis and machine intelligence, vol. 46, no. 12, pp. 8348–8364, 2024

    2024

  17. [17]

    LoRA: Low-rank adaptation of large language models,

    E. J. Hu, Y. Shen, P . Walliset al., “LoRA: Low-rank adaptation of large language models,”ICLR, vol. 1, no. 2, p. 3, 2022

    2022

  18. [18]

    Implicit style-content separation using b-lora,

    Y. Frenkel, Y. Vinker, A. Shamir, and D. Cohen-Or, “Implicit style-content separation using b-lora,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 181–198

    2024

  19. [19]

    Ziplora: Any subject in any style by effectively merging loras,

    V . Shah, N. Ruiz, F. Coleet al., “Ziplora: Any subject in any style by effectively merging loras,” inEuropean Conference on Computer Vision, 2024, pp. 422–438

    2024

  20. [20]

    Txt2img-mhn: Remote sensing image generation from text using modern hopfield networks,

    Y. Xu, W. Yu, P . Ghamisiet al., “Txt2img-mhn: Remote sensing image generation from text using modern hopfield networks,” IEEE T ransactions on Image Processing, vol. 32, pp. 5737–5750, 2023

    2023

  21. [21]

    Hopfield networks is all you need,

    H. Ramsauer, B. Sch ¨afl, J. Lehneret al., “Hopfield networks is all you need,”arXiv preprint arXiv:2008.02217, 2020

    Pith/arXiv arXiv 2008

  22. [22]

    Crs-diff: Controllable remote sensing image generation with diffusion model,

    D. Tang, X. Cao, X. Houet al., “Crs-diff: Controllable remote sensing image generation with diffusion model,”IEEE T ransactions on Geoscience and Remote Sensing, vol. 62, pp. 1–14, 2024

    2024

  23. [23]

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

    L. Zhang, A. Rao, and M. Agrawala, “Adding conditional control to text-to-image diffusion models,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 3836–3847

    2023

  24. [24]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P . Abbeel, “Denoising diffusion probabilistic models,”Advances in Neural Information Processing Systems, vol. 33, pp. 6840–6851, 2020

    2020

  25. [25]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,”arXiv preprint arXiv:2010.02502, 2020

    Pith/arXiv arXiv 2010

  26. [26]

    Hierarchical text- conditional image generation with clip latents,

    A. Ramesh, P . Dhariwal, A. Nicholet al., “Hierarchical text- conditional image generation with clip latents,”arXiv preprint arXiv:2204.06125, vol. 1, no. 2, p. 3, 2022

    Pith/arXiv arXiv 2022

  27. [27]

    Photorealistic text-to-image diffusion models with deep language understanding,

    C. Saharia, W. Chan, S. Saxenaet al., “Photorealistic text-to-image diffusion models with deep language understanding,”Advances in neural information processing systems, vol. 35, pp. 36 479–36 494, 2022

    2022

  28. [28]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agar- wal, G. Sastry, A. Askell, P . Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learning. PmLR, 2021, pp. 8748–8763

    2021

  29. [29]

    Unic-adapter: Unified image- instruction adapter with multi-modal transformer for image gen- eration,

    L. Duan, S. Zhao, W. Yanet al., “Unic-adapter: Unified image- instruction adapter with multi-modal transformer for image gen- eration,” inProceedings of the Computer Vision and Pattern Recogni- tion Conference, 2025, pp. 7963–7973

    2025

  30. [30]

    Alignguard: Scalable safety align- ment for text-to-image generation,

    R. Liu, I. C. Chen, J. Guet al., “Alignguard: Scalable safety align- ment for text-to-image generation,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 17 024–17 034

    2025

  31. [31]

    Enhancing mmdit-based text-to-image models for similar subject generation,

    T. Wei, D. Chen, Y. Zhou, and X. Pan, “Enhancing mmdit-based text-to-image models for similar subject generation,”IEEE T rans- actions on Pattern Analysis and Machine Intelligence, pp. 1–15, 2026

    2026

  32. [32]

    Cspanet: cross-route statistical partition attention network for style transfer,

    Y. Zhang, Y. Dong, and X. Pei, “Cspanet: cross-route statistical partition attention network for style transfer,”Neurocomputing, p. 131091, 2025

    2025

  33. [33]

    Disentangled representation learning,

    X. Wang, H. Chen, S. Tanget al., “Disentangled representation learning,”IEEE T ransactions on Pattern Analysis and Machine Intelli- gence, vol. 46, no. 12, pp. 9677–9696, 2024

    2024

  34. [34]

    Qr-lora: Efficient and disentangled fine-tuning via qr decomposition for customized generation,

    J. Yang, Y. Ma, D. Diet al., “Qr-lora: Efficient and disentangled fine-tuning via qr decomposition for customized generation,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 17 587–17 597

    2025

  35. [35]

    Auto-encoding variational bayes,

    D. P . Kingma and M. Welling, “Auto-encoding variational bayes,” arXiv preprint arXiv:1312.6114, 2013

    Pith/arXiv arXiv 2013

  36. [36]

    Singular value decomposi- tion for genome-wide expression data processing and modeling,

    O. Alter, P . O. Brown, and D. Botstein, “Singular value decomposi- tion for genome-wide expression data processing and modeling,” Proceedings of the National Academy of Sciences, vol. 97, no. 19, pp. 10 101–10 106, 2000

    2000

  37. [37]

    Adam: A method for stochastic optimiza- tion,

    D. P . Kingma and J. Ba, “Adam: A method for stochastic optimiza- tion,”arXiv preprint arXiv:1412.6980, 2014

    Pith/arXiv arXiv 2014

  38. [38]

    Mcanet: A joint semantic seg- mentation framework of optical and sar images for land use classification,

    X. Li, G. Zhang, H. Cuiet al., “Mcanet: A joint semantic seg- mentation framework of optical and sar images for land use classification,”International Journal of Applied Earth Observation and Geoinformation, vol. 106, p. 102638, 2022. 15

    2022

  39. [39]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochre- iter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,”Advances in neural information processing systems, vol. 30, 2017

    2017

  40. [40]

    Improved techniques for training gans,

    T. Salimans, I. Goodfellow, W. Zaremba, V . Cheung, A. Radford, and X. Chen, “Improved techniques for training gans,”Advances in neural information processing systems, vol. 29, 2016

    2016

  41. [41]

    Remote sensing image scene classification: Benchmark and state of the art,

    G. Cheng, J. Han, and X. Lu, “Remote sensing image scene classification: Benchmark and state of the art,”Proceedings of the IEEE, vol. 105, no. 10, pp. 1865–1883, 2017

    2017

  42. [42]

    Remote-sensing cross-domain scene classification: A dataset and benchmark,

    K. Liu, J. Yang, and S. Li, “Remote-sensing cross-domain scene classification: A dataset and benchmark,”Remote Sens., vol. 14, no. 18, p. 4635, 2022

    2022

  43. [43]

    Sardet-100k: Towards open-source bench- mark and toolkit for large-scale sar object detection,

    Y. Li, X. Li, W. Liet al., “Sardet-100k: Towards open-source bench- mark and toolkit for large-scale sar object detection,”Advances in Neural Information Processing Systems, vol. 37, pp. 128 430–128 461, 2024

    2024

  44. [44]

    Bright: A realistic and challeng- ing benchmark for reasoning-intensive retrieval,

    H. Su, H. Yen, M. Xiaet al., “Bright: A realistic and challeng- ing benchmark for reasoning-intensive retrieval,”arXiv preprint arXiv:2407.12883, 2024

    Pith/arXiv arXiv 2024

  45. [45]

    Fusar-ship: Building a high-resolution sar-ais matchup dataset of gaofen-3 for ship detection and recognition,

    X. Hou, W. Ao, Q. Song, J. Lai, H. Wang, and F. Xu, “Fusar-ship: Building a high-resolution sar-ais matchup dataset of gaofen-3 for ship detection and recognition,”Science China Information Sciences, vol. 63, no. 4, p. 140303, 2020

    2020

  46. [46]

    Drone-based rgb-infrared cross-modality vehicle detection,

    Y. Sun, B. Cao, P . Zhu, and Q. Hu, “Drone-based rgb-infrared cross-modality vehicle detection,”IEEE T rans. Circuits Syst. Video T echnol., vol. 32, no. 10, pp. 6700–6713, 2022

    2022

  47. [47]

    Vehicle detection in aerial imagery (vedai): a benchmark,

    S. Razakarivony and F. Jurie, “Vehicle detection in aerial imagery (vedai): a benchmark,” Tech. Rep., 2015. 2, Tech. Rep., 2014

    2015

  48. [48]

    Df- gan: A simple and effective baseline for text-to-image synthesis,

    M. Tao, H. Tang, F. Wu, X.-Y. Jing, B.-K. Bao, and C. Xu, “Df- gan: A simple and effective baseline for text-to-image synthesis,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 16 515–16 525

    2022

  49. [49]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,”arXiv preprint arXiv:1409.1556, 2014

    Pith/arXiv arXiv 2014

  50. [50]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProc. CVPR, 2016, pp. 770–778

    2016

This paper was first reviewed by grok-4.5 on July 30, 2026.