Pith. sign in

REVIEW 2 major objections 4 minor 77 references

AIComposer: Any Style and Content Image Composition via Feature Integration

T0 review · 2 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that style and content can be separated and recombined in CLIP image space, enabling prompt-free cross-domain image composition.

desk verdict A credible prompt-free cross-domain composition system whose central additive-CLIP claim needs a no-MLP baseline and a held-out test before the mechanism is proven, but the method and benchmark are worth referee time. read the letter →

arxiv 2507.20721 v1 pith:H3RENNP4 submitted 2025-07-28 cs.CV

Haowen Li , Zhenfeng Fan , Zhang Wen , Zhengzhou Zhu , Yunjin Li This is my paper
classification cs.CV
keywords cross-domainimagecompositionprompt-freeCLIPfeatureintegrationdiffusionmodelsstyletransferIP-Adapterresidualblendingmaskedcross-attention
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 cross-domain image composition can be done without any text prompt by operating entirely on CLIP image features (the image embeddings of the Contrastive Language-Image Pre-training model). The proposed method, AIComposer, takes the CLIP features of a foreground object and a background scene, blends them with a small multilayer perceptron through the residual identity $f_{\text{integrate}} = f_c + f_s - F(f_c, f_s)$, and injects the result into a latent diffusion model through masked cross-attention. If this works as claimed, it is the first prompt-free system that preserves foreground content while transferring background style, and it reports large gains over prior work: LPIPS reduced 30.5% and CSD up 18.1% on the paper's benchmarks. The practical stake is that many styles and subjects are hard to describe in words, so a method that learns the blend from images rather than prompts could make composition tools usable in far more settings.

What carries the argument

The load-bearing object is the residual feature-integration identity $f_{\text{integrate}} = f_c + f_s - F(f_c, f_s)$: the CLIP features of the foreground content and background style are added, and a learned three-layer MLP $F$ supplies a residual correction trained by $\mathcal{L} = \sum_k \|F(f_c^k, f_s^k) - (f_c^k + f_s^k - f_l^k)\|_2^2$. The paper motivates this with a linear discriminant analysis showing that content and style classes cluster in CLIP image feature space, and it leans on the additive property of CLIP as a robustness anchor so that even a poorly trained residual still leaves the input content and style features present. This integrated feature is then injected into the Stable Diffusion XL denoising process through masked cross-attention with an additive mask and adaptive instance normalization, applied only in the first few steps and only in the dilated foreground region.

What would settle it

Take held-out content-style pairs, compute the integrated feature $f_c + f_s - F(f_c, f_s)$, and measure its cosine similarity to the IP-Adapter feature of a human-verified stylized version of the same pair; if the similarity is no better than using $f_c$ or $f_s$ alone, the additive CLIP assumption is not doing the work.

Watch

Extended reading notes

Core claim

The paper's central claim is that style and content are separable and additively recombinable in the CLIP image feature space used by the IP-Adapter, an image-prompt adapter that injects CLIP image features into a diffusion model, so an integrated feature can be obtained as $f_{\text{integrate}} = f_c + f_s - F(f_c, f_s)$, where $F$ is a three-layer MLP trained on content-style-stylized triplets. The paper argues that this feature, injected only inside a dilated foreground mask through rectified cross-attention, turns a cross-domain composition into an effectively same-domain one: the diffusion model harmonizes the stylized foreground with the background while a few adaptive-instance-normalization operations align the tone. It reports that this design beats Blended Latent Diffusion, Paint by Example, AnyDoor, and TF-ICON on both the TF-ICON benchmark and a new extended benchmark, with LPIPS reduced 30.5% and CSD improved 18.1%, and that it works even with no text prompt at all.

Load-bearing premise

The whole mechanism rests on the assumption that, for images the MLP has never seen, the style and content of a CLIP feature can be separated and recombined by simple addition, with the learned residual only cleaning up the blend.

Editorial extensions

If this is right

  • A user can compose a foreground and background even when the desired style or subject has no easy verbal description; a short prompt such as 'a fox' suffices and an empty prompt still works.
  • Because the method inverts only the blended latent and runs ten diffusion steps in a single branch, it is cheaper than two-branch full-step composition methods.
  • Since the diffusion backbone is not modified, the approach combines with ControlNet edge conditioning and can swap the roles of foreground and background to stylize either image.
  • The released extended benchmark gives cross-domain composition a common evaluation ground, with backgrounds spanning sketches, watercolors, oil paintings, cartoon animation, sci-fi, and pixel art.

Reading between the lines

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

  • If the additive decomposability of CLIP image features holds beyond the tested pairs, the same residual identity could be reused for other feature-space edits, such as prompt-free object replacement, attribute transfer, or subject swapping, without retraining the diffusion model.
  • The residual learning design suggests that the MLP's exact output matters less than the input features, so the method may tolerate much smaller triplet sets; this is testable by training on a few hundred pairs with the same residual loss.
  • The success of ten-step single-branch inversion raises the possibility that the same prompt-free composition idea transfers to video or interactive editing, where per-frame cost matters, though the paper does not evaluate those settings.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper proposes AIComposer, a prompt-free cross-domain image composition method built on Stable Diffusion XL (SDXL) and IP-Adapter. A foreground and a background image are blended in pixel space and then in latent space via AdaIN, inverted with DPM-Solver++, and denoised while injecting a blended CLIP feature into masked cross-attention. The blended feature is produced by a trained 3-layer MLP that integrates foreground content with background style; the MLP is trained in a residual manner on content-style-stylized triplets. The authors introduce an extended cross-domain benchmark and report improvements over prior methods in LPIPS, CSD, PSNR, and CLIP-T, supported by a user study.

Significance. If the central mechanism—that content and style are additively separable in CLIP image feature space—holds, this would be a useful and practical contribution: prompt-free cross-domain composition with a small MLP and a single-branch diffusion process. The paper is commendable for releasing code and a new benchmark, for running component ablations, and for including a user study, which provides independent grounding beyond the automated metrics. However, the evidence for the load-bearing additive compositionality assumption is thin (an LDA visualization), and the headline LPIPS metric is in tension with the paper's own ablation results, weakening the quantitative superiority claim. The work is likely of interest to the image-composition community, but its central claims need stronger support before publication.

major comments (2)
  1. [§4.3, Tables 1–3] The LPIPS metric is computed against the naive hybrid paste (foreground pasted onto background) within the masked region, so lower LPIPS values reward outputs that remain close to the unstylized paste. The paper's own ablation in Table 3 shows that removing the initial latent blending (-InitBlend) improves LPIPS from 0.4195 to 0.3376 while substantially decreasing CSD from 0.5284 to 0.4671, confirming that LPIPS rewards less stylization. Although the paper acknowledges metric limitations in §4.4, the abstract and introduction headline the 30.5% LPIPS improvement. This makes the quantitative claim state-of-the-art results load-bearing on a metric that the authors themselves show is inversely related to style transfer. Please re-evaluate the headline claims by reporting LPIPS against a stylized reference or by de-emphasizing LPIPS relative to style-consistency metrics.
  2. [§3.3, Eq. (1), Fig. 3, Table 4] The central assertion is that CLIP image features are additively separable into content and style, so that f_integrate = f_c + f_s - F(f_c, f_s) approximates the feature of a stylized foreground. The evidence for this assertion is an LDA visualization on 80 images (Fig. 3), which demonstrates linear class separability in two projected dimensions, not additive compositionality in the original feature space. No held-out triplet test is reported comparing the integrated feature (with or without the learned residual) against ground-truth stylized features, and the no-MLP baseline f_integrate = f_c + f_s is never evaluated. The small performance difference between 300 and 30,000 training triplets in Table 4 is consistent with the MLP contributing only slightly beyond the additive shortcut. Without a no-MLP baseline and a quantitative test of additivity, the paper's claim that a small MLP can separate and integrate content and style in CLIP space is not supported.
minor comments (4)
  1. [§3.3] The phrase "additive property of CLIP [40]" is not established by the cited Radford et al. paper; please provide a demonstration or a more specific citation that supports additive compositionality of CLIP image features.
  2. [Contributions, §3.3] The contribution bullet states that "style and content are linearly separable," but the proposed integration uses a nonlinear MLP and the LDA experiment shows separability after a learned linear projection, not in the original feature space; the wording should be adjusted.
  3. [Table 1 vs. Table 2] The CLIP-T computation uses a global image-to-text comparison for TF-ICON samples and a masked-region-to-text comparison for the extended benchmark, as described in Sec. 12; please indicate which variant is used in each table, since the numbers are not directly comparable.
  4. [Supplementary, Fig. 13] The MLP code uses input_dim = 2*2048*8 and output_dim = 2048*8; the text does not specify the exact CLIP feature dimension before IP-Adapter projection. Clarify this to ensure reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the MLP is a transparent supervised fit, and the additive-CLIP premise is an empirical assumption rather than a self-referential derivation.

full rationale

The derivation chain in Sec. 3.3 is not circular by the paper's own equations. Eq. 1 trains F to regress on externally generated CSGO triplets with target f_c+f_s-f_l, and inference uses f_c+f_s-F(f_c,f_s); the algebraic relation between the training target and the inference formula means that a perfectly trained F would reproduce f_l on the training distribution, but this is standard residual supervised learning, not a claim that an unseen stylized feature is derived from the inputs without labels. Evaluation is performed on TF-ICON and a newly constructed benchmark with a user study, so the comparisons are not statistically forced by the training objective. The weak point is the asserted 'additive property of CLIP [40]', which is an empirical premise supported only by an LDA cluster plot on 80 images; cluster separability does not establish additive compositionality, and the paper itself concedes that content and style cannot be thoroughly decoupled. That is a correctness and evidence concern, not a circularity. No load-bearing self-citation, imported uniqueness theorem, or ansatz-by-self-citation appears; the references to CSGO and IP-Adapter are external and used as training-data and source-feature providers. The residual-robustness argument (if F is random, f_c+f_s still contributes) is a stated design rationale, not a derivation that assumes its conclusion. Overall the paper is self-contained against external benchmarks, and the central mechanism, while possibly under-supported, is not equivalent to its inputs by construction.

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

The method contributes a trained MLP and an evaluation benchmark; the derived-looking content/style integration is actually a supervised regression onto CSGO style-transfer features. Except for standard network design, the paper adds no new mathematical axioms, so the ledger is dominated by domain assumptions about CLIP feature structure.

free parameters (5)
  • MLP weights (25.45M parameters) = trained on N=30,000 triplets, Adam lr=1e-4, hidden_dim=512
    The central style-content integration F(f_c, f_s) is a learned network; its weights are fit to CSGO-generated triplets, not derived.
  • AdaIN blending ratios lambda = lambda=1 for initial latent blend; lambda=0.1 during diffusion
    Set by hand; controls color-tone stylization strength.
  • Diffusion manipulation window = first 5 denoising steps
    Chosen without ablation; later steps run free.
  • Inversion and reconstruction step count = 10 steps
    Selected by parameter impact analysis; 5, 15, and 20 steps trade rigidity versus variation.
  • Training-data filtering thresholds = 65,429 triplets reduced to 37,445 manually
    Subjective exclusion criteria (content inconsistency, style discrepancy, visual quality) determine the learned mapping.
assumptions (5)
  • domain assumption Content and style are linearly separable in CLIP image feature space.
    Supported by LDA clustering on 80 images per class (Sec. 3.3), not proven; Eq. 1 and the residual inference scheme rely on this decomposition.
  • domain assumption CLIP image features are additive: f_stylized is roughly f_content plus f_style minus a residual.
    Invoked in Sec. 3.3 as the additive property of CLIP; used to justify robustness even when the MLP residual is poorly learned.
  • domain assumption IP-Adapter CLIP image features can steer SDXL cross-attention to transfer content and style.
    Borrowed from IP-Adapter and used throughout the pipeline; no verification specific to this composition task.
  • domain assumption DPM-Solver++ inversion of the blended latent for 10 steps preserves sufficient content and style information.
    Inversion is a numerical approximation; the paper relies on it to reconstruct the initial blend (Sec. 3.2) without a formal error bound.
  • domain assumption The frozen SDXL diffusion prior suffices for arbitrary cross-domain compositions.
    The method deliberately avoids diffuser training, so the success of stylization depends on the pretrained prior for all unseen style-content pairs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AIComposer: Any Style and Content Image Composition via Feature Integration." pith.science (2026). https://pith.science/paper/H3RENNP4

@misc{pith2026250720721,
  author       = {Pith},
  title        = {Pith review of: AIComposer: Any Style and Content Image Composition via Feature Integration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H3RENNP4}},
  note         = {Machine review of arXiv:2507.20721}
}
read the original abstract

Image composition has advanced significantly with large-scale pre-trained T2I diffusion models. Despite progress in same-domain composition, cross-domain composition remains under-explored. The main challenges are the stochastic nature of diffusion models and the style gap between input images, leading to failures and artifacts. Additionally, heavy reliance on text prompts limits practical applications. This paper presents the first cross-domain image composition method that does not require text prompts, allowing natural stylization and seamless compositions. Our method is efficient and robust, preserving the diffusion prior, as it involves minor steps for backward inversion and forward denoising without training the diffuser. Our method also uses a simple multilayer perceptron network to integrate CLIP features from foreground and background, manipulating diffusion with a local cross-attention strategy. It effectively preserves foreground content while enabling stable stylization without a pre-stylization network. Finally, we create a benchmark dataset with diverse contents and styles for fair evaluation, addressing the lack of testing datasets for cross-domain image composition. Our method outperforms state-of-the-art techniques in both qualitative and quantitative evaluations, significantly improving the LPIPS score by 30.5% and the CSD metric by 18.1%. We believe our method will advance future research and applications. Code and benchmark at https://github.com/sherlhw/AIComposer.

Figures

Figures reproduced from arXiv: 2507.20721 by the authors.

Figure 1
Figure 1. Comparison between our proposed method and a representative state-of-the-art approach TF-ICON [ [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overall architecture including the (a) pipeline, the (b) image blending strategy, and the (c) MLP training process. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Clustering results of the first two dimensions after apply [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (11 more)
Figure 5
Figure 5. Figure 5: Qualitative comparison of our method with prior SOTA works. We also include the decoded image of the initial blended latent [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 7
Figure 7. Figure 7: Parameter Impact Analysis: (Top) different diffusion [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 6
Figure 6. Figure 6: The average user preference percentage for different [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 8
Figure 8. Figure 8: Two examples of the composited results and their [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Examples of our composited results (the 2nd column), and blended feature visualization (the last 6 columns) using different number of training data, with and without residual training strategy. 9. Results for same-domain composition Despite its powerful capabilities in…
Figure 10
Figure 10. Figure 10: The qualitative results of ablation study. [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 12
Figure 12. Figure 12: An example of exchanging the roles of the foreground [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]
Figure 13
Figure 13. Figure 13: Codes for the 3-layer MLP network. Principles of Filtering Training Data for MLP. We gen￾erate content-style stylized triplets by leveraging a state-of￾the-art neural-style transfer method [65], resulting in a pre￾liminary collection of 65, 429 triplets. The quality a…
Figure 14
Figure 14. Figure 14: Qualitative comparison of our method with prior SOTA works. We also include an early version of stable diffusion (Ours [PITH_FULL_IMAGE:figures/full_fig_p014_14.png]
Figure 15
Figure 15. Figure 15: Qualitative comparison of our method with prior SOTA works. We also include an early version of stable diffusion (Ours [PITH_FULL_IMAGE:figures/full_fig_p015_15.png]
Figure 16
Figure 16. Figure 16: Qualitative comparison of our method with prior SOTA works. We also include an early version of stable diffusion (Ours [PITH_FULL_IMAGE:figures/full_fig_p016_16.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

77 extracted references · 61 canonical work pages

  1. [1]

    Blended diffusion for text-driven editing of natural images

    Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, pages 18187–18197, 2022. 3

    work page 2022

  2. [2]

    Blended latent diffusion

    Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Trans. Graph., 42(4):149:1–149:11,

    work page

  3. [3]

    Kim, Zhaowen Wang, Eli Shechtman, and Trevor Darrell

    Samaneh Azadi, Matthew Fisher, Vladimir G. Kim, Zhaowen Wang, Eli Shechtman, and Trevor Darrell. Multi- content GAN for few-shot font style transfer. InCVPR, pages 7564–7573, 2018. 3

    work page 2018

  4. [4]

    John F. Canny. A computational approach to edge detection. IEEE TPAMI, 8(6):679–698, 1986. 3

    work page 1986

  5. [5]

    Toward realistic im- age compositing with adversarial learning

    Bor-Chun Chen and Andrew Kae. Toward realistic im- age compositing with adversarial learning. In CVPR, pages 8415–8424, 2019. 3

    work page 2019

  6. [6]

    Anydoor: Zero-shot object-level im- age customization

    Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level im- age customization. In CVPR, pages 6593–6602, 2024. 3, 6, 7, 2

    work page 2024

  7. [7]

    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. In ECCV, 2024. 3

    work page 2024

  8. [8]

    Style injec- tion in diffusion: A training-free approach for adapting large- scale diffusion models for style transfer

    Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injec- tion in diffusion: A training-free approach for adapting large- scale diffusion models for style transfer. In CVPR, pages 8795–8805, 2024. 1

    work page 2024

Show all 77 references
  1. [9]

    The use of multiple measurements in tax- onomic problems

    Ronald A Fisher. The use of multiple measurements in tax- onomic problems. Annals of eugenics, 7(2):179–188, 1936. 5

    1936

  2. [10]

    Implicit style-content separation using b-lora

    Yarden Frenkel, Yael Vinker, Ariel Shamir, and Daniel Cohen-Or. Implicit style-content separation using b-lora. In ECCV, 2024. 3

    2024

  3. [11]

    An image is worth one word: Personalizing text-to-image generation using textual inversion

    Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 2

    2023

  4. [12]

    Gatys, Alexander S

    Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. Texture synthesis using convolutional neural networks. In NeurIPS, pages 262–270, 2015. 3

    2015

  5. [13]

    Gatys, Alexander S

    Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. A neural algorithm of artistic style. Journal of Vision, 16(12): 326, 2016. 3

    2016

  6. [14]

    Caltech-256 object category dataset, 2023

    Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset, 2023. 5

    2023

  7. [15]

    a cartoon animation of a goose in the forest

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium. NeurIPS, 30, 2017. 2 4 Input TF-ICON Ours-SD1.5Blended Diffusion Paint-by-Example Anydoor Ours-NP Our...

    2017

  8. [16]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. In NeurIPS, pages 6840–6851,

  9. [17]

    Stylemix: Sep- arating content and style for enhanced data augmentation

    Minui Hong, Jinwoo Choi, and Gunhee Kim. Stylemix: Sep- arating content and style for enhanced data augmentation. In CVPR, pages 14862–14870, 2021. 3

    2021

  10. [18]

    Belongie

    Xun Huang and Serge J. Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, pages 1510–1519, 2017. 3, 4

    2017

  11. [19]

    Designedit: Multi- layered latent decomposition and fusion for unified & accu- rate image editing

    Yueru Jia, Yuhui Yuan, Aosong Cheng, Chuke Wang, Ji Li, Huizhu Jia, and Shanghang Zhang. Designedit: Multi- layered latent decomposition and fusion for unified & accu- rate image editing. CoRR, abs/2403.14487, 2024. 3

    2024 arXiv

  12. [20]

    Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion

    Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In ECCV,

  13. [21]

    Imagic: Text-based real image editing with diffusion models

    Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In CVPR, pages 6007–6017, 2023. 2

    2023

  14. [22]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6

    2015

  15. [23]

    Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross B

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chlo´e Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C. Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross B. Girshick. Segment anything. In ICCV, pages 3992– 4003, 2023. 6

    2023

  16. [24]

    Multi-concept customiza- tion of text-to-image diffusion

    Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customiza- tion of text-to-image diffusion. In CVPR, pages 1931–1941,

    1931

  17. [25]

    Dongxu Li, Junnan Li, and Steven C. H. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to- image generation and editing. In NeurIPS, 2023. 2

    2023

  18. [26]

    Layerdiffusion: Layered controlled image editing with dif- fusion models

    Pengzhi Li, Qinxuan Huang, Yikang Ding, and Zhiheng Li. Layerdiffusion: Layered controlled image editing with dif- fusion models. In SIGGRAPH, pages 12:1–12:4, 2023. 2

    2023

  19. [27]

    Dreamedit: Subject-driven image editing

    Tianle Li, Max Ku, Cong Wei, and Wenhu Chen. Dreamedit: Subject-driven image editing. Trans. Mach. Learn. Res. ,

  20. [28]

    Style- tokenizer: Defining image style by a single instance for con- trolling diffusion models

    Wen Li, Muyuan Fang, Cheng Zou, Biao Gong, Ruobing Zheng, Meng Wang, Jingdong Chen, and Ming Yang. Style- tokenizer: Defining image style by a single instance for con- trolling diffusion models. CoRR, abs/2409.02543, 2024. 6

    2024 arXiv

  21. [29]

    Motrans: Customized motion transfer with text-driven video diffusion models

    Xiaomin Li, Xu Jia, Qinghe Wang, Haiwen Diao, Meng- meng Ge, Pengxiang Li, You He, and Huchuan Lu. Motrans: Customized motion transfer with text-driven video diffusion models. In ACM MM, pages 3421–3430, 2024. 2

    2024

  22. [30]

    ST-GAN: spatial transformer generative adversarial networks for image compositing

    Chen-Hsuan Lin, Ersin Yumer, Oliver Wang, Eli Shechtman, and Simon Lucey. ST-GAN: spatial transformer generative adversarial networks for image compositing. InCVPR, pages 9455–9464, 2018. 3

    2018

  23. [31]

    Dpm-solver: A fast ODE solver for dif- fusion probabilistic model sampling in around 10 steps

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ODE solver for dif- fusion probabilistic model sampling in around 10 steps. In NeurIPS, pages 5775–5787, 2022. 3, 6

    2022

  24. [32]

    Dreamcom: Finetuning text-guided inpainting model for image composition

    Lingxiao Lu, Bo Zhang, and Li Niu. Dreamcom: Finetuning text-guided inpainting model for image composition. CoRR, abs/2309.15508, 2023. 3

    2023 arXiv

  25. [33]

    TF- ICON: diffusion-based training-free cross-domain image composition

    Shilin Lu, Yanzhu Liu, and Adams Wai-Kin Kong. TF- ICON: diffusion-based training-free cross-domain image composition. In ICCV, pages 2294–2305, 2023. 1, 2, 3, 4, 6, 7

    2023

  26. [34]

    Repaint: Inpainting using denoising diffusion probabilistic models

    Andreas Lugmayr, Martin Danelljan, Andr´es Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11451–11461, 2022. 2

    2022

  27. [35]

    Null-text inversion for editing real images using guided diffusion models

    Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, pages 6038–6047,

  28. [36]

    Making images real again: A comprehensive survey on deep image composition

    Li Niu, Wenyan Cong, Liu Liu, Yan Hong, Bo Zhang, Jing Liang, and Liqing Zhang. Making images real again: A comprehensive survey on deep image composition. CoRR, abs/2106.14490, 2021. 3

    2021

  29. [37]

    Poisson image editing

    Patrick P ´erez, Michel Gangnet, and Andrew Blake. Poisson image editing. ACM TOG, 22(3):313–318, 2003. 3

    2003

  30. [38]

    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 optimization. CoRR, abs/2408.03637, 2024. 2, 3, 4, 6

    2024 arXiv

  31. [39]

    SDXL: improving latent diffusion mod- els for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion mod- els for high-resolution image synthesis. In ICLR, 2024. 3, 2

    2024

  32. [40]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748–...

    2021

  33. [41]

    Hierarchical text-conditional image gener- ation with CLIP latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener- ation with CLIP latents. CoRR, abs/2204.06125, 2022. 2

    2022 arXiv

  34. [42]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. In CVPR, pages 10674– 10685, 2022. 2, 3

    2022

  35. [43]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500–22510, 2023. 2, 6

    2023

  36. [44]

    Jacobs, and Shlomi Fruchter

    Nataniel Ruiz, Yuanzhen Li, Neal Wadhwa, Yael Pritch, Michael Rubinstein, David E. Jacobs, and Shlomi Fruchter. Magic insert: Style-aware drag-and-drop. CoRR, abs/2407.02489, 2024. 3, 6

    2024 arXiv

  37. [45]

    Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Moham- mad Norouzi. Photorealistic text-to-image diffusion mod- els wit...

    2022

  38. [46]

    Elgammal

    Babak Saleh and Ahmed M. Elgammal. Large-scale classifi- cation of fine-art paintings: Learning the right metric on the right feature. Journal of Computer Science & Technology , 25(3):595–605, 2015. 5

    2015

  39. [47]

    Ziplora: Any subject in any style by effectively merging loras

    Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuanzhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. In ECCV, pages 422–438, 2025. 2

    2025

  40. [48]

    Hierarchical image saliency detection on extended CSSD

    Jianping Shi, Qiong Yan, Li Xu, and Jiaya Jia. Hierarchical image saliency detection on extended CSSD. IEEE TPAMI, 38(4):717–729, 2016. 6

    2016

  41. [49]

    Weiss, Niru Mah- eswaranathan, and Surya Ganguli

    Jascha Sohl-Dickstein, Eric A. Weiss, Niru Mah- eswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256–2265, 2015. 2

    2015

  42. [50]

    Measuring style similarity in diffusion models

    Gowthami Somepalli, Anubhav Gupta, Kamal Gupta, Shra- may Palta, Micah Goldblum, Jonas Geiping, Abhinav Shri- vastava, and Tom Goldstein. Measuring style similarity in diffusion models. CoRR, abs/2404.01292, 2024. 7

    2024 arXiv

  43. [51]

    Generative modeling by es- timating gradients of the data distribution

    Yang Song and Stefano Ermon. Generative modeling by es- timating gradients of the data distribution. InNeurIPS, pages 11895–11907, 2019. 2

    2019

  44. [52]

    Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equa- tions. In ICLR, 2021. 2

    2021

  45. [53]

    Lin, Scott Cohen, Brian L

    Yizhi Song, Zhifei Zhang, Zhe L. Lin, Scott Cohen, Brian L. Price, Jianming Zhang, Soo Ye Kim, and Daniel G. Aliaga. Objectstitch: Object compositing with diffusion model. In CVPR, pages 18310–18319, 2023. 2, 3

    2023

  46. [54]

    Price, Jianming Zhang, Soo Ye Kim, He Zhang, Wei Xiong, and Daniel G

    Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian L. Price, Jianming Zhang, Soo Ye Kim, He Zhang, Wei Xiong, and Daniel G. Aliaga. IMPRINT: generative object com- positing by learning identity-preserving representation. In CVPR, pages 8048–8058, 2024. 6

    2024

  47. [55]

    Thinking out- side the bbox: Unconstrained generative object compositing

    Gemma Canet Tarres, Zhe Lin, Zhifei Zhang, Jianming Zhang, Yizhi Song, Dan Sebastian Ruta, Andrew Gilbert, John Philip Collomosse, and Soo Ye Kim. Thinking out- side the bbox: Unconstrained generative object compositing. In ECCV, pages 476–495, 2024. 3

    2024

  48. [56]

    Instantstyle: Free lunch to- wards style-preserving in text-to-image generation

    Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch to- wards style-preserving in text-to-image generation. CoRR, abs/2404.02733, 2024. 2, 3, 4

    2024 arXiv

  49. [57]

    Multimodality-guided image style transfer using cross-modal GAN inversion

    Hanyu Wang, Pengxiang Wu, Kevin Dela Rosa, Chen Wang, and Abhinav Shrivastava. Multimodality-guided image style transfer using cross-modal GAN inversion. In WACV, pages 4964–4973, 2024. 3

    2024

  50. [58]

    Instantstyle-plus: Style transfer with content-preserving in text-to-image generation

    Haofan Wang, Peng Xing, Renyuan Huang, Hao Ai, Qixun Wang, and Xu Bai. Instantstyle-plus: Style transfer with content-preserving in text-to-image generation. CoRR, abs/2407.00788, 2024. 6

    2024 arXiv

  51. [59]

    Salient object detection: A discriminative regional feature integration ap- proach

    Jingdong Wang, Huaizu Jiang, Zejian Yuan, Ming-Ming Cheng, Xiaowei Hu, and Nanning Zheng. Salient object detection: A discriminative regional feature integration ap- proach. Int. J. Comput. Vis., 123(2):251–268, 2017. 6

    2017

  52. [60]

    Primecomposer: Faster progressively combined diffu- sion for image composition with attention steering

    Yibin Wang, Weizhong Zhang, Jianwei Zheng, and Cheng Jin. Primecomposer: Faster progressively combined diffu- sion for image composition with attention steering. In ACM MM, pages 10824–10832, 2024. 2, 3, 4, 5, 6

    2024

  53. [61]

    Elite: Encoding visual con- cepts into textual embeddings for customized text-to-image generation

    Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual con- cepts into textual embeddings for customized text-to-image generation. In ICCV, pages 15943–15953, 2023. 2

    2023

  54. [62]

    Artfid: Quantitative evaluation of neural style transfer

    Matthias Wright and Bj ¨orn Ommer. Artfid: Quantitative evaluation of neural style transfer. In DAGM German Con- ference on Pattern Recognition, pages 560–576, 2022. 2

    2022

  55. [63]

    Fastcomposer: Tuning-free multi- subject image generation with localized attention

    Guangxuan Xiao, Tianwei Yin, William T Freeman, Fr ´edo Durand, and Song Han. Fastcomposer: Tuning-free multi- subject image generation with localized attention. IJCV, pages 1–20, 2024. 2

    2024

  56. [64]

    Smartbrush: Text and shape guided object inpainting with diffusion model

    Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, pages 22428–22437, 2023. 3

    2023

  57. [65]

    CSGO: content-style composition in text-to-image genera- tion

    Peng Xing, Haofan Wang, Yanpeng Sun, Qixun Wang, Xu Bai, Hao Ai, Renyuan Huang, and Zechao Li. CSGO: content-style composition in text-to-image genera- tion. CoRR, abs/2408.16766, 2024. 3, 5, 6, 1, 4

    2024 arXiv

  58. [66]

    Freetuner: Any subject in any style with training-free diffu- sion

    Youcan Xu, Zhen Wang, Jun Xiao, Wei Liu, and Long Chen. Freetuner: Any subject in any style with training-free diffu- sion. CoRR, abs/2405.14201, 2024. 4

    2024 arXiv

  59. [67]

    Paint by example: Exemplar-based image editing with diffusion mod- els

    Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion mod- els. In CVPR, pages 18381–18391, 2023. 2, 3, 6, 7

    2023

  60. [68]

    Pastiche master: Exemplar-based high-resolution por- trait style transfer

    Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Pastiche master: Exemplar-based high-resolution por- trait style transfer. In CVPR, pages 7683–7692, 2022. 3

    2022

  61. [69]

    Ip- adapter: Text compatible image prompt adapter for text-to- image diffusion models

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip- adapter: Text compatible image prompt adapter for text-to- image diffusion models. CoRR, abs/2308.06721, 2023. 2, 3, 4, 5, 1

    2023 arXiv

  62. [70]

    Style trans- fer for anime sketches with enhanced residual u-net and aux- iliary classifier GAN

    Lvmin Zhang, Yi Ji, Xin Lin, and Chunping Liu. Style trans- fer for anime sketches with enhanced residual u-net and aux- iliary classifier GAN. In ACPR, pages 506–511, 2017. 3

    2017

  63. [71]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3813–3824, 2023. 2, 3

    2023

  64. [72]

    Efros, Eli Shecht- man, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586– 595, 2018. 7, 2

    2018

  65. [73]

    Separating style and content for generalized style transfer

    Yexun Zhang, Ya Zhang, and Wenbin Cai. Separating style and content for generalized style transfer. In CVPR, pages 8447–8455, 2018. 3

    2018

  66. [74]

    A unified frame- work for generalizable style transfer: Style and content sep- aration

    Yexun Zhang, Ya Zhang, and Wenbin Cai. A unified frame- work for generalizable style transfer: Style and content sep- aration. IEEE TIP, 29:4085–4098, 2020. 3

    2020

  67. [75]

    CFA-GAN: cross fusion attention and frequency loss for image style transfer

    Xiangtian Zheng, Xiaolin Yang, Qun Zhao, Hao Zhang, Xi- ufeng He, Jinsong Zhang, and Xuan Zhang. CFA-GAN: cross fusion attention and frequency loss for image style transfer. Displays, 81:102588, 2024. 3 9

    2024

  68. [76]

    Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle- consistent adversarial networks. In ICCV, pages 2242–2251,

  69. [77]

    A task is worth one word: Learning with task prompts for high-quality versatile image inpainting

    Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. In ECCV, 2024. 2 10

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.