HarmoniDiff-RS: Training-Free Diffusion Harmonization for Satellite Image Composition

Jefersson A. dos Santos; Jungong Han; Xiaoqi Zhuang

arxiv: 2604.19392 · v1 · submitted 2026-04-21 · 💻 cs.CV

HarmoniDiff-RS: Training-Free Diffusion Harmonization for Satellite Image Composition

Xiaoqi Zhuang , Jefersson A. dos Santos , Jungong Han This is my paper

Pith reviewed 2026-05-10 03:32 UTC · model grok-4.3

classification 💻 cs.CV

keywords satellite image harmonizationdiffusion modelstraining-free methodimage compositionremote sensinglatent fusiondomain alignmentbenchmark dataset

0 comments

The pith

A training-free diffusion framework harmonizes composite satellite images by aligning domains in latent space.

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

The paper establishes that satellite image composites from different domains can be made visually coherent without training any new models by shifting radiometric properties through latent operations and fusing information across diffusion timesteps. This matters for remote sensing because applications such as data augmentation, disaster simulation, and urban planning require large numbers of realistic combined images that preserve original content while matching in appearance. The approach generates multiple candidate composites and uses a lightweight classifier to pick the most consistent result. If the claim holds, it removes the need for domain-specific retraining and supports scalable synthesis of satellite data under varying conditions.

Core claim

HarmoniDiff-RS performs satellite image composition by applying a Latent Mean Shift operation to transfer radiometric characteristics between source and target domains, followed by a Timestep-wise Latent Fusion strategy that draws on early inverted latents for high harmonization and late latents for semantic consistency to produce candidate results, and finally selects the best output using a trained harmony classifier. The method is validated on the newly constructed RSIC-H benchmark containing 500 paired composition samples derived from fMoW, with experiments showing effective performance for remote-sensing synthesis and simulation tasks.

What carries the argument

The Timestep-wise Latent Fusion strategy, which combines early inverted latents for domain harmonization with late latents for content preservation, supported by Latent Mean Shift for radiometric alignment and a harmony classifier for final selection.

If this is right

Composite satellite images become available for data augmentation without requiring domain-specific model retraining.
Disaster simulation and urban planning tasks gain access to domain-consistent imagery generated on demand.
The harmony classifier provides an automatic quality filter that scales with the number of generated candidates.
Remote sensing workflows can synthesize training data across multiple source domains in a single pass.

Where Pith is reading between the lines

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

The latent fusion approach could extend to harmonizing non-satellite imagery if the early-to-late timestep principle holds in other diffusion setups.
It might reduce reliance on large paired training sets for remote sensing models by enabling synthetic composite creation.
Adapting the mean shift and fusion steps to handle temporal sequences could support video-based satellite applications.
Testing the method on more extreme domain gaps, such as cross-sensor or cross-season pairs, would clarify its robustness limits.

Load-bearing premise

The Timestep-wise Latent Fusion strategy using early inverted latents for harmonization and late latents for semantic consistency reliably produces coherent composites across diverse domain conditions without introducing artifacts that the harmony classifier cannot detect.

What would settle it

Visible seams, radiometric mismatches, or semantic distortions appearing in the selected composites when tested on satellite image pairs with large differences in lighting, sensor type, or geography would show that the fusion and selection process fails to deliver reliable harmonization.

Figures

Figures reproduced from arXiv: 2604.19392 by Jefersson A. dos Santos, Jungong Han, Xiaoqi Zhuang.

**Figure 1.** Figure 1: Comparison between natural [12] (upper) and satellite (bottom) image composition. Natural image composition emphasizes semantic alignment and may involve non-rigid source deformation, while satellite composition focuses on boundary harmonization under geometric rigidity. changing the pose or shape to better fit the surrounding context (e.g., the sheep in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Figure 2: Overview of the proposed HarmoniDiff-RS framework. Given an initial composition of a target scene and a source patch, the model first applies Latent Mean Shift to align the source latent with the target style. Then, Timestep-wise Latent Sampling generates a sequence of intermediate compositions: early timesteps yield more harmonious but less faithful results, while late timesteps better preserve structural… view at source ↗

**Figure 3.** Figure 3: Effect of Latent Mean Shift (LMS) under different vari [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Figure 4: Qualitative comparison on RSIC-H. Our method generates seamless and semantically aligned compositions, outperforming [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

**Figure 6.** Figure 6: From left to right: (1) severe semantic and contextual [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

**Figure 5.** Figure 5: The target image (left) undergoes significant high [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

read the original abstract

Satellite image composition plays a critical role in remote sensing applications such as data augmentation, disaste simulation, and urban planning. We propose HarmoniDiff-RS, a training-free diffusion-based framework for harmonizing composite satellite images under diverse domain conditions. Our method aligns the source and target domains through a Latent Mean Shift operation that transfers radiometric characteristics between them. To balance harmonization and content preservation, we introduce a Timestep-wise Latent Fusion strategy by leveraging early inverted latents for high harmonization and late latents for semantic consistency to generate a set of composite candidates. A lightweight harmony classifier is trained to further automatically select the most coherent result among them. We also construct RSIC-H, a benchmark dataset for satellite image harmonization derived from fMoW, providing 500 paired composition samples. Experiments demonstrate that our method effectively performs satellite image composition, showing strong potential for scalable remote-sensing synthesis and simulation tasks. Code is available at: https://github.com/XiaoqiZhuang/HarmoniDiff-RS.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

HarmoniDiff-RS gives a training-free diffusion route to harmonize satellite composites plus a new 500-pair benchmark, but the strength of the gains rests on results that the abstract does not show.

read the letter

The paper puts forward HarmoniDiff-RS, a diffusion pipeline that avoids any training step for satellite image harmonization. It shifts latent means to transfer radiometric traits, then fuses early and late inverted latents at different timesteps to trade off domain match against content preservation, and finally uses a small trained classifier to pick the best candidate from the set. They also release RSIC-H, a 500-sample paired benchmark built from fMoW, and make the code public on GitHub. That combination is the concrete addition here: an operational recipe tailored to remote-sensing composites and a test set that did not exist before for this exact task. People working on data augmentation or simulation in satellite imagery can grab the code and the pairs without starting from scratch. The training-free design is a practical plus when paired data is hard to collect at scale. The soft spots sit in the validation. The abstract states that experiments demonstrate effectiveness, yet supplies no numbers, no baseline comparisons, and no ablation on the fusion choices or the classifier. Without those details it is difficult to judge whether the timestep split actually separates domain alignment from semantics cleanly across sensor variations, or whether the classifier simply masks residual seams and spectral mismatches. The stress-test concern about artifacts slipping past the selector therefore remains open until the full results table is examined. This work is aimed at researchers who already handle satellite composites for augmentation or planning tasks. A reader in applied remote-sensing CV will find the dataset and the released implementation directly usable. It is not a broad theoretical step, so general computer-vision groups will likely skip it. The paper deserves a serious referee because the new benchmark and the code lower the barrier for others to test and extend the idea, even if the evaluation section will probably need more baselines and error analysis before publication.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes HarmoniDiff-RS, a training-free diffusion-based framework for harmonizing composite satellite images. It aligns domains via a Latent Mean Shift operation that transfers radiometric characteristics, employs a Timestep-wise Latent Fusion strategy that combines early inverted latents (for harmonization) with late latents (for semantic consistency) to produce candidate composites, and uses a separately trained lightweight harmony classifier to select the most coherent output. The authors introduce the RSIC-H benchmark of 500 paired composition samples derived from fMoW and claim that experiments demonstrate effective satellite image composition with potential for scalable remote-sensing synthesis tasks.

Significance. If the central claims are substantiated, the work would offer a practical training-free alternative for satellite image harmonization that avoids costly retraining of diffusion models, supporting data augmentation, disaster simulation, and urban planning applications. The release of the RSIC-H benchmark is a constructive addition to the remote-sensing community. The significance is currently limited by insufficient empirical grounding of the fusion mechanism's ability to separate domain alignment from semantic preservation across sensor variations.

major comments (3)

[§3.3] §3.3 (Timestep-wise Latent Fusion): The claim that early inverted latents reliably supply domain alignment while late latents preserve semantics is load-bearing for the effectiveness argument, yet the manuscript provides no ablation studies on the timestep split point, no quantitative metrics (e.g., boundary seam detection or spectral mismatch scores) comparing fused outputs to non-fused baselines, and no analysis of failure modes where the harmony classifier selects composites containing undetected artifacts.
[§4] §4 (Experiments): The reported results on the 500-sample RSIC-H benchmark lack direct comparisons to established harmonization baselines (both traditional and learning-based) and do not include cross-domain stress tests or error analysis showing that the classifier reliably detects inconsistencies arising from radiometric and textural variations typical in satellite imagery.
[§4.1] §4.1 (Benchmark): The RSIC-H dataset size of 500 paired samples may not sufficiently cover the diversity of sensor-specific domain shifts; the manuscript does not demonstrate that performance generalizes beyond the fMoW-derived pairs or quantify how often the fusion step introduces subtle inconsistencies missed by the classifier.

minor comments (2)

[Abstract] Abstract: 'disaste simulation' is a typographical error and should read 'disaster simulation'.
[§3.1] §3.1: The mathematical description of the Latent Mean Shift operation would benefit from explicit pseudocode or a numbered equation to improve reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We appreciate the emphasis on strengthening the empirical validation of the Timestep-wise Latent Fusion and the experimental sections. Below we provide point-by-point responses to the major comments, indicating revisions we will make to address the concerns while preserving the core contributions of the training-free framework and the RSIC-H benchmark.

read point-by-point responses

Referee: [§3.3] §3.3 (Timestep-wise Latent Fusion): The claim that early inverted latents reliably supply domain alignment while late latents preserve semantics is load-bearing for the effectiveness argument, yet the manuscript provides no ablation studies on the timestep split point, no quantitative metrics (e.g., boundary seam detection or spectral mismatch scores) comparing fused outputs to non-fused baselines, and no analysis of failure modes where the harmony classifier selects composites containing undetected artifacts.

Authors: We agree that additional validation is required to support the design choices in Timestep-wise Latent Fusion. In the revised manuscript we will add ablation experiments varying the split point (e.g., early timesteps t=100–300 versus later t=500–700) and report quantitative metrics including boundary seam detection via gradient consistency scores, spectral mismatch via histogram KL divergence, and semantic preservation via feature similarity from a pretrained encoder. We will also include a dedicated failure-mode analysis section that examines cases where the harmony classifier selects outputs with residual artifacts, supported by visual examples and discussion of when the early/late latent assumption breaks down. These results will be added to §3.3 and the supplementary material. revision: yes
Referee: [§4] §4 (Experiments): The reported results on the 500-sample RSIC-H benchmark lack direct comparisons to established harmonization baselines (both traditional and learning-based) and do not include cross-domain stress tests or error analysis showing that the classifier reliably detects inconsistencies arising from radiometric and textural variations typical in satellite imagery.

Authors: We acknowledge the value of broader benchmarking. The revised Section 4 will incorporate direct quantitative comparisons against traditional baselines (histogram matching, Poisson blending, and seamless cloning) and representative learning-based harmonization methods, using the same RSIC-H pairs and standard metrics (PSNR, SSIM, LPIPS, and a seam-visibility score). We will further add cross-domain stress tests on held-out sensor pairs (e.g., mixing Landsat and Sentinel-2 imagery) and provide error analysis of the harmony classifier, including its precision-recall on detecting radiometric and textural inconsistencies. These additions will directly address the referee’s concern about empirical grounding. revision: yes
Referee: [§4.1] §4.1 (Benchmark): The RSIC-H dataset size of 500 paired samples may not sufficiently cover the diversity of sensor-specific domain shifts; the manuscript does not demonstrate that performance generalizes beyond the fMoW-derived pairs or quantify how often the fusion step introduces subtle inconsistencies missed by the classifier.

Authors: The 500-pair RSIC-H benchmark is presented as an initial, publicly released resource derived from fMoW to enable standardized evaluation; we agree that broader sensor coverage would be beneficial. In revision we will expand §4.1 with a limitations discussion, additional qualitative results on non-fMoW imagery where available, and a quantitative error study that manually inspects a random subset of classifier-selected outputs to measure the frequency and nature of subtle inconsistencies introduced by fusion. While a substantially larger multi-sensor corpus lies beyond the current scope, we will frame the existing benchmark size transparently and list dataset expansion as future work. revision: partial

Circularity Check

0 steps flagged

No circularity: method steps are independent operations, not reductions to inputs

full rationale

The paper proposes concrete algorithmic steps (Latent Mean Shift for domain alignment, Timestep-wise Latent Fusion using early/late inverted latents, and a separately trained lightweight harmony classifier for selection) whose definitions and execution do not reduce by construction to the target outputs or to self-citations. The RSIC-H benchmark is constructed from fMoW but is an external data resource, not a fitted parameter. No equations or claims equate a 'prediction' to a fitted input, smuggle an ansatz via self-citation, or rename a known result as a derivation. Central claims rest on empirical experiments rather than tautological equivalence, making the framework self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The approach rests on standard diffusion model assumptions about latent space behavior and the utility of timestep-specific information; no new physical entities are postulated and free parameters appear limited to the lightweight classifier training.

axioms (2)

domain assumption Diffusion models allow controllable image editing through latent manipulation at different timesteps
Invoked to justify the mean shift and fusion operations for harmonization versus content preservation
domain assumption A lightweight classifier can reliably identify the most coherent composite among generated candidates
Basis for the automatic selection step

pith-pipeline@v0.9.0 · 5479 in / 1373 out tokens · 45819 ms · 2026-05-10T03:32:20.325126+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages · 1 internal anchor

[1]

Freecompose: Generic zero-shot image composition with diffusion prior

Zhekai Chen, Wen Wang, Zhen Yang, Zeqing Yuan, Hao Chen, and Chunhua Shen. Freecompose: Generic zero-shot image composition with diffusion prior. arXiv preprint arXiv:2407.04947, 2024. 1, 2, 3, 6

work page arXiv 2024
[2]

Functional map of the world

Gordon Christie, Neil Fendley, James Wilson, and Ryan Mukherjee. Functional map of the world. In CVPR, 2018. 4, 5

work page 2018
[3]

Deep residual learning for image recognition

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5

work page 2016
[4]

Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in neural infor- mation processing systems, 30, 2017

Martin Heusel, Hubert Ramsauer, Thomas Un- terthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in neural infor- mation processing systems, 30, 2017. 5, 6

work page 2017
[5]

Denois- ing diffusion probabilistic models

Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denois- ing diffusion probabilistic models. InAdvances in Neural Information Processing Systems, pages 6840–

work page
[6]

Curran Associates, Inc., 2020. 2, 3

work page 2020
[7]

Lobell, and Stefano Ermon

Samar Khanna, Patrick Liu, Linqi Zhou, Chenlin Meng, Robin Rombach, Marshall Burke, David B. Lobell, and Stefano Ermon. Diffusionsat: A gener- ative foundation model for satellite imagery. InThe Twelfth International Conference on Learning Repre- sentations, 2024. 2, 5, 8

work page 2024
[8]

Auto-Encoding Variational Bayes

Diederik P Kingma and Max Welling. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114,

work page internal anchor Pith review Pith/arXiv arXiv
[9]

Text2earth: Unlocking text-driven remote sensing image generation with a global-scale dataset and a foundation model.IEEE Geoscience and Remote Sensing Magazine, pages 2–23, 2025

Chenyang Liu, Keyan Chen, Rui Zhao, Zhengxia Zou, and Zhenwei Shi. Text2earth: Unlocking text-driven remote sensing image generation with a global-scale dataset and a foundation model.IEEE Geoscience and Remote Sensing Magazine, pages 2–23, 2025. 2

work page 2025
[10]

Tf- icon: Diffusion-based training-free cross-domain im- age composition

Shilin Lu, Yanzhu Liu, and Adams Wai-Kin Kong. Tf- icon: Diffusion-based training-free cross-domain im- age composition. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 2294–2305, 2023. 1, 3

work page 2023
[11]

Tkg-dm: Training-free chroma key content generation diffusion model.arXiv preprint arXiv:2411.15580, 2024

Ryugo Morita, Stanislav Frolov, Brian Bernhard Moser, Takahiro Shirakawa, Ko Watanabe, Andreas Dengel, and Jinjia Zhou. Tkg-dm: Training-free chroma key content generation diffusion model.arXiv preprint arXiv:2411.15580, 2024. 3

work page arXiv 2024
[12]

Poisson image editing.ACM Transactions on Graph- ics, 22(3):313–318, 2003

Patrick P ´erez, Michel Gangnet, and Andrew Blake. Poisson image editing.ACM Transactions on Graph- ics, 22(3):313–318, 2003. 1, 2, 6

work page 2003
[13]

Pham, Jingye Chen, and Qifeng Chen

Kien T. Pham, Jingye Chen, and Qifeng Chen. TALE: Training-free cross-domain image composition via adaptive latent manipulation and energy-guided opti- mization. InACM Multimedia 2024, 2024. 1, 2, 3

work page 2024
[14]

High- resolution image synthesis with latent diffusion mod- els, 2021

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High- resolution image synthesis with latent diffusion mod- els, 2021. 3, 5, 6, 8

work page 2021
[15]

Geosynth: Contextually-aware high- resolution satellite image synthesis

Srikumar Sastry, Subash Khanal, Aayush Dhakal, and Nathan Jacobs. Geosynth: Contextually-aware high- resolution satellite image synthesis. InIEEE/ISPRS Workshop: Large Scale Computer Vision for Remote Sensing (EARTHVISION), 2024. 2

work page 2024
[16]

Denoising diffusion implicit models

Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. InInternational Conference on Learning Representations, 2021. 2, 3

work page 2021
[17]

Deep image harmonization in dual color spaces.arXiv preprint arXiv:2308.02813, 2023

Linfeng Tan, Jiangtong Li, Li Niu, and Liqing Zhang. Deep image harmonization in dual color spaces.arXiv preprint arXiv:2308.02813, 2023. 2

work page arXiv 2023
[18]

Crs- diff: Controllable remote sensing image generation with diffusion model.IEEE Transactions on Geo- science and Remote Sensing, 2024

Datao Tang, Xiangyong Cao, Xingsong Hou, Zhongyuan Jiang, Junmin Liu, and Deyu Meng. Crs- diff: Controllable remote sensing image generation with diffusion model.IEEE Transactions on Geo- science and Remote Sensing, 2024. 2

work page 2024
[19]

Aero- gen: Enhancing remote sensing object detection with diffusion-driven data generation.arXiv preprint arXiv:2411.15497, 2024

Datao Tang, Xiangyong Cao, Xuan Wu, Jialin Li, Jing Yao, Xueru Bai, and Deyu Meng. Aero- gen: Enhancing remote sensing object detection with diffusion-driven data generation.arXiv preprint arXiv:2411.15497, 2024. 2

work page arXiv 2024
[20]

arXiv preprint arXiv:2211.13227 , year=

Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based im- age editing with diffusion models.arXiv preprint arXiv:2211.13227, 2022. 2

work page arXiv 2022
[21]

Deep image blending

Lingzhi Zhang, Tarmily Wen, and Jianbo Shi. Deep image blending. InThe IEEE Winter Conference on Applications of Computer Vision, pages 231–240,

work page
[22]

Cc-diff++: Spatially controllable text- to-image synthesis for remote sensing with enhanced contextual coherence.IEEE Transactions on Geo- science and Remote Sensing, 63:1–16, 2025

Mu Zhang, Yunfan Liu, Yue Liu, Yuzhong Zhao, and Qixiang Ye. Cc-diff++: Spatially controllable text- to-image synthesis for remote sensing with enhanced contextual coherence.IEEE Transactions on Geo- science and Remote Sensing, 63:1–16, 2025. 2

work page 2025

[1] [1]

Freecompose: Generic zero-shot image composition with diffusion prior

Zhekai Chen, Wen Wang, Zhen Yang, Zeqing Yuan, Hao Chen, and Chunhua Shen. Freecompose: Generic zero-shot image composition with diffusion prior. arXiv preprint arXiv:2407.04947, 2024. 1, 2, 3, 6

work page arXiv 2024

[2] [2]

Functional map of the world

Gordon Christie, Neil Fendley, James Wilson, and Ryan Mukherjee. Functional map of the world. In CVPR, 2018. 4, 5

work page 2018

[3] [3]

Deep residual learning for image recognition

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5

work page 2016

[4] [4]

Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in neural infor- mation processing systems, 30, 2017

Martin Heusel, Hubert Ramsauer, Thomas Un- terthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in neural infor- mation processing systems, 30, 2017. 5, 6

work page 2017

[5] [5]

Denois- ing diffusion probabilistic models

Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denois- ing diffusion probabilistic models. InAdvances in Neural Information Processing Systems, pages 6840–

work page

[6] [6]

Curran Associates, Inc., 2020. 2, 3

work page 2020

[7] [7]

Lobell, and Stefano Ermon

Samar Khanna, Patrick Liu, Linqi Zhou, Chenlin Meng, Robin Rombach, Marshall Burke, David B. Lobell, and Stefano Ermon. Diffusionsat: A gener- ative foundation model for satellite imagery. InThe Twelfth International Conference on Learning Repre- sentations, 2024. 2, 5, 8

work page 2024

[8] [8]

Auto-Encoding Variational Bayes

Diederik P Kingma and Max Welling. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114,

work page internal anchor Pith review Pith/arXiv arXiv

[9] [9]

Text2earth: Unlocking text-driven remote sensing image generation with a global-scale dataset and a foundation model.IEEE Geoscience and Remote Sensing Magazine, pages 2–23, 2025

Chenyang Liu, Keyan Chen, Rui Zhao, Zhengxia Zou, and Zhenwei Shi. Text2earth: Unlocking text-driven remote sensing image generation with a global-scale dataset and a foundation model.IEEE Geoscience and Remote Sensing Magazine, pages 2–23, 2025. 2

work page 2025

[10] [10]

Tf- icon: Diffusion-based training-free cross-domain im- age composition

Shilin Lu, Yanzhu Liu, and Adams Wai-Kin Kong. Tf- icon: Diffusion-based training-free cross-domain im- age composition. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 2294–2305, 2023. 1, 3

work page 2023

[11] [11]

Tkg-dm: Training-free chroma key content generation diffusion model.arXiv preprint arXiv:2411.15580, 2024

Ryugo Morita, Stanislav Frolov, Brian Bernhard Moser, Takahiro Shirakawa, Ko Watanabe, Andreas Dengel, and Jinjia Zhou. Tkg-dm: Training-free chroma key content generation diffusion model.arXiv preprint arXiv:2411.15580, 2024. 3

work page arXiv 2024

[12] [12]

Poisson image editing.ACM Transactions on Graph- ics, 22(3):313–318, 2003

Patrick P ´erez, Michel Gangnet, and Andrew Blake. Poisson image editing.ACM Transactions on Graph- ics, 22(3):313–318, 2003. 1, 2, 6

work page 2003

[13] [13]

Pham, Jingye Chen, and Qifeng Chen

Kien T. Pham, Jingye Chen, and Qifeng Chen. TALE: Training-free cross-domain image composition via adaptive latent manipulation and energy-guided opti- mization. InACM Multimedia 2024, 2024. 1, 2, 3

work page 2024

[14] [14]

High- resolution image synthesis with latent diffusion mod- els, 2021

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High- resolution image synthesis with latent diffusion mod- els, 2021. 3, 5, 6, 8

work page 2021

[15] [15]

Geosynth: Contextually-aware high- resolution satellite image synthesis

Srikumar Sastry, Subash Khanal, Aayush Dhakal, and Nathan Jacobs. Geosynth: Contextually-aware high- resolution satellite image synthesis. InIEEE/ISPRS Workshop: Large Scale Computer Vision for Remote Sensing (EARTHVISION), 2024. 2

work page 2024

[16] [16]

Denoising diffusion implicit models

Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. InInternational Conference on Learning Representations, 2021. 2, 3

work page 2021

[17] [17]

Deep image harmonization in dual color spaces.arXiv preprint arXiv:2308.02813, 2023

Linfeng Tan, Jiangtong Li, Li Niu, and Liqing Zhang. Deep image harmonization in dual color spaces.arXiv preprint arXiv:2308.02813, 2023. 2

work page arXiv 2023

[18] [18]

Crs- diff: Controllable remote sensing image generation with diffusion model.IEEE Transactions on Geo- science and Remote Sensing, 2024

Datao Tang, Xiangyong Cao, Xingsong Hou, Zhongyuan Jiang, Junmin Liu, and Deyu Meng. Crs- diff: Controllable remote sensing image generation with diffusion model.IEEE Transactions on Geo- science and Remote Sensing, 2024. 2

work page 2024

[19] [19]

Aero- gen: Enhancing remote sensing object detection with diffusion-driven data generation.arXiv preprint arXiv:2411.15497, 2024

Datao Tang, Xiangyong Cao, Xuan Wu, Jialin Li, Jing Yao, Xueru Bai, and Deyu Meng. Aero- gen: Enhancing remote sensing object detection with diffusion-driven data generation.arXiv preprint arXiv:2411.15497, 2024. 2

work page arXiv 2024

[20] [20]

arXiv preprint arXiv:2211.13227 , year=

Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based im- age editing with diffusion models.arXiv preprint arXiv:2211.13227, 2022. 2

work page arXiv 2022

[21] [21]

Deep image blending

Lingzhi Zhang, Tarmily Wen, and Jianbo Shi. Deep image blending. InThe IEEE Winter Conference on Applications of Computer Vision, pages 231–240,

work page

[22] [22]

Cc-diff++: Spatially controllable text- to-image synthesis for remote sensing with enhanced contextual coherence.IEEE Transactions on Geo- science and Remote Sensing, 63:1–16, 2025

Mu Zhang, Yunfan Liu, Yue Liu, Yuzhong Zhao, and Qixiang Ye. Cc-diff++: Spatially controllable text- to-image synthesis for remote sensing with enhanced contextual coherence.IEEE Transactions on Geo- science and Remote Sensing, 63:1–16, 2025. 2

work page 2025