Pith. sign in

REVIEW 4 major objections 6 minor 54 references

Diffusion model separates object from style for image retrieval

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 diffusion-based disentanglement method that separates object content from domain style achieves state-of-the-art unsupervised cross-domain image retrieval on three benchmarks.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection DUDE is a serious UCIR paper with a genuinely new diffusion-based disentanglement idea and large SOTA gains, but the load-bearing disentanglement claim needs independent verification before I'd fully trust the mechanism. the 4 major comments →

arxiv 2509.04193 v1 pith:WQZF5XQ7 submitted 2025-09-04 cs.CV cs.LG

DUDE: Diffusion-Based Unsupervised Cross-Domain Image Retrieval

Ruohong Yang , Peng Hu , Yunfan Li , Xi Peng This is my paper
classification cs.CV cs.LG
keywords unsupervised cross-domain image retrievalfeature disentanglementdiffusion modelsStable Diffusionobject-semantic tokenprogressive alignmentcontrastive learningimage retrieval
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

The paper tries to establish that unsupervised cross-domain image retrieval can be solved far more accurately by first disentangling domain-invariant object features from domain-specific styles, using a frozen text-to-image diffusion model as the disentangler, and only then aligning those clean features across domains. It claims that a per-image object-semantic token, trained by minimizing the diffusion denoising loss while the domain-style token is fixed in the prompt, learns only clean object features. On three standard benchmarks covering 13 domain pairs, DUDE beats prior unsupervised retrieval methods by large margins, e.g., average P@50 of 71.09 on PACS versus 50.26 for the best previous method. If correct, the work would show that generative diffusion models can serve as feature disentanglers for retrieval, and that a disentangle-then-align paradigm is more effective than aligning entangled whole-image features.

Core claim

DUDE's central claim is that the domain gap in unsupervised cross-domain image retrieval is best attacked by explicitly separating object content from style before any cross-domain alignment, and that a pretrained text-to-image diffusion model can perform that separation in token space. For each image, the feature extractor produces an object-semantic token [z], which is inserted into a prompt along with a handcrafted domain token ("a sketch of a [z]"). Freezing all Stable Diffusion parameters and back-propagating only into the feature extractor through the denoising objective forces [z] to encode the image content that the diffusion model cannot otherwise reproduce when the style is already

What carries the argument

The load-bearing mechanism is the Object Disentanglement Module built on a frozen Stable Diffusion. The prompt "a {domain} of a [z]" pairs a fixed style token with a learnable per-image object token; minimizing the diffusion denoising loss (Eq. 4) with respect to the feature extractor only, while all diffusion parameters stay frozen, is what forces [z] to absorb object semantics and leave style to the domain token. The second mechanism is the Progressive Alignment Module, which uses an InfoNCE-based contrastive loss with mutual top-k neighbor filtering, first in-domain, then cross-domain, supported by a momentum encoder and per-domain memory banks from MoCo.

Load-bearing premise

The key assumption is that minimizing the diffusion denoising loss with a fixed style token forces the per-image token to encode only clean object semantics and not residual style or low-level image details; the paper asserts this in Section 3.2 but does not independently measure disentanglement.

What would settle it

A concrete test: within a single object category, train a linear classifier to predict the source domain from DUDE's [z] embeddings; accuracy well above chance means style information remains in the object token and the disentanglement claim is incomplete. A second test: freeze [z], vary the domain token, generate images, and check with a pretrained object classifier whether the object identity survives the style swap consistently across all domain pairs.

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

If this is right

  • Cross-domain retrieval systems could use object-only embeddings, making matches robust to changes in sketch, painting, or photographic style.
  • The disentanglement objective could be reused for other cross-domain tasks such as classification, person re-identification, or sketch-based image retrieval.
  • Stable Diffusion's token space can be mined for discriminative object features even when its generative text-to-image capabilities are not the goal.
  • The progressive alignment schedule (instance → in-domain → cross-domain) could reduce the need for large batches or costly pairwise mining in other unsupervised retrieval settings.

Where Pith is reading between the lines

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

  • The claimed disentanglement is directly testable: a linear classifier on [z] trying to predict the domain label should stay near chance level within a category; if the domain is easily read off from [z], style still leaks into the object token.
  • The method's ceiling likely depends on how well the pretrained diffusion model covers the target styles; for unseen or strongly non-photographic styles, the fixed domain token may not fully absorb the style, and the approach would need additional style tokens or fine-tuning.
  • If the disentanglement is as clean as claimed, the same [z] could enable zero-shot style transfer at inference time by simply swapping the domain token, which the paper's qualitative generation examples already hint at.
  • The progressive alignment idea could be combined with any shared-embedding objective, not just retrieval, whenever within-domain structure is more reliable than cross-domain structure at the start of training.
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 / 6 minor

Summary. The paper proposes DUDE, an unsupervised cross-domain image retrieval (UCIR) method that combines a Stable-Diffusion-based object disentanglement module with a progressive alignment module. For each image, a learnable object-semantic token [z] is inserted into a domain-specific prompt (e.g., 'a sketch of a [z]'); the feature extractor is trained by minimizing the diffusion denoising loss with a frozen Stable Diffusion model, with the intention that [z] captures domain-invariant object content while the fixed domain token absorbs style. The resulting clean object features are then aligned progressively: augmented views, in-domain mutual nearest neighbors, and finally cross-domain mutual nearest neighbors. Experiments on PACS, Office-Home, and DomainNet report large improvements over existing UCIR baselines, e.g., average P@50 of 71.09 vs. 50.26 for ProtoOT on PACS, and 82.62 vs. 66.63 on DomainNet. The paper includes ablations, parameter sensitivity analyses, retrieval visualizations, and t-SNE plots, and promises code release.

Significance. If the central disentanglement claim holds, the paper would introduce a genuinely new mechanism for UCIR and an interesting new use of text-to-image diffusion models beyond generation. The reported gains are large and consistent across 13 domains, and the proposed pipeline is clearly described and reproducible in principle. The paper also ships an ablation of the loss components and a prompt-style control. However, the load-bearing assertion that the token [z] learns only clean object features is not independently validated; the current evidence is largely the retrieval numbers themselves plus a few qualitative generation examples. Because the comparisons do not include a baseline with the same diffusion backbone but without the proposed disentanglement, the attribution of the gains to feature disentanglement remains unproven. The significance is therefore conditional on additional targeted experiments.

major comments (4)
  1. [Sec. 3.2, Eq. (4)] The paper repeatedly states that the object-semantic token [z] 'learns only the clean object features' because the domain-style token is fixed in the prompt. This is not established by the objective in Eq. (4). The diffusion denoising loss requires [z] to provide whatever information is needed to predict the noise for the specific input image, which includes pose, background, color, and potentially residual style details. The fixed domain token captures only a coarse, handcrafted style label (e.g., 'sketch'), not the full distribution of stylistic variation within a domain. The claim therefore needs an independent test. Concretely, I would expect a domain-classification probe on the learned [z] features, a quantitative style-transfer fidelity metric for Fig. 3, or a learned-token reconstruction experiment comparing the information retained in [z] versus a randomly initialized token. With
  2. [Tab. 5] The fixed-prompt control ('a photo of a [z]' for all domains) is the right idea, but its results partially undermine the style-disentanglement narrative. On the four DomainNet pairs shown, the average P@50 gap between varying and fixed prompts is only about 5 points (67.46 vs. 62.33, and 67.89 vs. 64.31 in the second block), and on some individual rows the gap is much smaller (e.g., Painting→Clipart: 91.01 vs. 90.65). This indicates that the domain-style token is not the dominant source of the retrieval gains; a fixed prompt still gives near-SOTA performance. The authors should either perform a significance test over multiple runs, or add a control where the prompt is varied randomly but not matched to the true domain. The current results do not exclude the possibility that the benefit comes from prompt diversity or from the diffusion loss acting as a strong regularizer, rather than from
  3. [Sec. 4.5, Tab. 6] The comparison to baselines is not apples-to-apples. All baselines use a ResNet-50 initialized from ImageNet and are trained with instance/cluster contrastive losses, whereas DUDE additionally uses a frozen Stable Diffusion v1.5 as a large pretrained generative backbone. The ablation row 'ResNet50' in Tab. 6 controls for neither the additional pretrained capacity nor the auxiliary diffusion objective. To support the claim that disentanglement—not just the diffusion regularizer—is responsible for the gains, the authors should include a control that uses the same diffusion-loss objective with a randomly initialized token, or with a generic diffusion-based representation loss, but without the domain-style prompt. This would isolate the source of the improvement and make the SOTA claim more convincing.
  4. [Sec. 4.7, Fig. 4] The parameter analysis is not described as being performed on a validation split. The statement 'Based on these observations, we set k=50, β=0.5, and λ=1 for all experiments' suggests that these values were chosen after inspecting performance on the same test domains. If so, the reported numbers are optimistic and the 'robustness' conclusion is circular. The authors should either describe a validation-based selection protocol or perform a nested cross-validation. At minimum, they should report the sensitivity of the final results to these choices (e.g., the full range of P@50 across the tested k, β, and λ values) and clarify whether any test set was used to select hyperparameters.
minor comments (6)
  1. [References] References [14] and [15] are identical (Huang et al., ICCV 2015), and references [44] and [45] are also identical (Wang et al., AAAI 2023). Please consolidate or correct these entries.
  2. [Sec. 4.7 heading] The heading reads 'Analysis on neighbors k' — should be 'Analysis of the number of neighbors k'.
  3. [Funding section] There is a typo: 'Finding No funding was received' should be 'Funding: No funding was received'.
  4. [Tab. 6] The last column header 'P@50(1) P@100(5) P@200(15)' is confusing because the numbers in parentheses seem to refer to the different metric families (P@50 for PACS, P@1 for Office-Home, etc.). Please clarify the notation.
  5. [Fig. 3] The qualitative generation result shows that SD can generate images in different styles from a learned token, but this does not by itself prove that the token is free of style information, since SD's style control could override residual style in the token. This point should be acknowledged and ideally backed by a quantitative measure.
  6. [Sec. 4.2] The paper does not discuss the computational cost of using Stable Diffusion v1.5 for every training sample, nor how this compares with the baselines. Given the substantial performance gains, a runtime/memory comparison would help readers assess practical impact.

Circularity Check

0 steps flagged

No significant circularity: DUDE's retrieval results are evaluated on held-out benchmarks and its disentanglement claim, while under-verified, is an empirical assumption rather than a definitional or self-citational reduction.

full rationale

The paper's central derivation chain is not circular in the sense defined here. The object-semantic token [z] is optimized by the diffusion denoising loss (Eq. 4) and thereafter used for retrieval; however, the retrieval metrics are computed on standard held-out test splits (PACS, Office-Home, DomainNet) and are not fitted or predicted from the training objective by construction. The claim that [z] learns 'only the clean object features' is asserted from an analogy with text-to-image generation, but it is not derived from an equation that equates the output with the input; it is an unverified modeling assumption, which is a correctness/robustness concern rather than a circularity. The paper includes ablations (e.g., varying vs. fixed domain prompts in Table 5, loss-component ablations in Table 6) that compare alternatives empirically rather than importing the conclusion through self-citation. There is no self-citation chain carrying the load: the cited Stable Diffusion, MoCo, and ProtoOT are external works, and no 'uniqueness theorem' from the authors is invoked to forbid alternatives. The qualitative cross-domain generation results are internal evidence but do not make the derivation tautological. Consequently, no specific circular step can be quoted with an exhibited reduction, and the appropriate finding is no significant circularity (score 0).

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The method rests on several unverified premises about Stable Diffusion's latent token space and on hand-tuned hyperparameters. The most important is that the diffusion denoising gradient pushes [z] to be object-only. The free parameters (k, beta, lambda, tau) are chosen by parameter sweeps without a stated validation protocol, which adds risk of selection bias.

free parameters (4)
  • k (number of mutual neighbors) = 50
    Chosen via parameter analysis (Fig. 4a); controls the number of neighbors for in-domain and cross-domain alignment.
  • beta (L_PA1 coefficient) = 0.5
    Weight of in-domain alignment loss in L_PA1 (Eq. 11), tuned via parameter analysis (Fig. 4b).
  • lambda (L_PA2 coefficient) = 1
    Weight of cross-domain alignment loss in L_PA2 (Eq. 17), tuned via parameter analysis (Fig. 4c).
  • tau (temperature) = 0.2
    Temperature parameter in InfoNCE losses; set 'across all datasets' per Sec. 4.2, standard in contrastive learning.
axioms (4)
  • domain assumption Stable Diffusion v1.5 pretrained on LAION is capable of separating object semantics from style in its token space.
    Sec. 3.2: the method assumes a text-to-image model that can synthesize consistent objects under different stylistic prompts implicitly encodes separable object and style tokens.
  • ad hoc to paper Backpropagating the diffusion denoising loss into the image-derived token [z] will compress domain-invariant object content rather than residual style or image-specific details.
    Eq. 4 in Sec. 3.2: this is the core training objective; the paper provides no theoretical or independent empirical proof that this gradient direction enforces disentanglement.
  • domain assumption Mutual top-k nearest neighbors in feature space are reliable proxies for same-class pairs, both within and across domains.
    Sec. 3.3, Eqs. 6-7, 12-13: standard but unverified assumption underlying the progressive alignment losses.
  • domain assumption The categories and label sets are shared across domains, and domain identity is known for every training image.
    Problem formulation in Sec. 3.1; standard for UCIR.
invented entities (1)
  • Object-semantic token [z_i] no independent evidence
    purpose: A per-image embedding, output by fθ and inserted into the text prompt, intended to capture domain-invariant object semantics.
    The token is introduced by the paper and trained via LOD; its disentanglement is only supported by retrieval performance and qualitative cross-domain generation in the paper, not by an external benchmark.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of DUDE: Diffusion-Based Unsupervised Cross-Domain Image Retrieval." pith.science (2026). https://pith.science/paper/WQZF5XQ7

@misc{pith2026250904193,
  author       = {Pith},
  title        = {Pith review of: DUDE: Diffusion-Based Unsupervised Cross-Domain Image Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WQZF5XQ7}},
  note         = {Machine review of arXiv:2509.04193}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Unsupervised cross-domain image retrieval (UCIR) aims to retrieve images of the same category across diverse domains without relying on annotations. Existing UCIR methods, which align cross-domain features for the entire image, often struggle with the domain gap, as the object features critical for retrieval are frequently entangled with domain-specific styles. To address this challenge, we propose DUDE, a novel UCIR method building upon feature disentanglement. In brief, DUDE leverages a text-to-image generative model to disentangle object features from domain-specific styles, thus facilitating semantical image retrieval. To further achieve reliable alignment of the disentangled object features, DUDE aligns mutual neighbors from within domains to across domains in a progressive manner. Extensive experiments demonstrate that DUDE achieves state-of-the-art performance across three benchmark datasets over 13 domains. The code will be released.

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

54 extracted references · 43 canonical work pages · 2 internal anchors

  1. [1]

    Segdiff: Image segmen- tation with diffusion probabilistic models

    Tomer Amit, Tal Shaharbany, Eliya Nach- mani, and Lior Wolf. Segdiff: Image segmen- tation with diffusion probabilistic models. arXiv preprint arXiv:2112.00390 , 2021

    Pith/arXiv arXiv 2021

  2. [2]

    Mmfl-net: multi-scale and multi-granularity feature learning for cross- domain fashion retrieval

    Chen Bao, Xudong Zhang, Jiazhou Chen, and Yongwei Miao. Mmfl-net: multi-scale and multi-granularity feature learning for cross- domain fashion retrieval. Multimedia Tools and Applications, 82(24):37905–37937, 2023

    work page 2023

  3. [3]

    FAST: Improving Controllability for Text Generation with Feedback Aware Self-Training

    Junyi Chai, Reid Pryzant, Victor Ye Dong, Konstantin Golobokov, Chenguang Zhu, and Yi Liu. Fast: Improving controllability for text generation with feedback aware self- training. arXiv preprint arXiv:2210.03167 , 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  4. [4]

    Data-free sketch-based image retrieval

    Abhra Chaudhuri, Ayan Kumar Bhunia, Yi- Zhe Song, and Anjan Dutta. Data-free sketch-based image retrieval. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition , pages 12084– 12093, 2023

    work page 2023

  5. [5]

    Dif- fusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Dif- fusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780–8794, 2021. 13

    work page 2021

  6. [6]

    Tutorial on variational autoen- coders

    Carl Doersch. Tutorial on variational autoen- coders. arXiv preprint arXiv:1606.05908 , 2016

    Pith/arXiv arXiv 2016

  7. [7]

    Cross- domain fashion image retrieval

    Bojana Gajic and Ramon Baldrich. Cross- domain fashion image retrieval. In Proceed- ings of the IEEE conference on computer vision and pattern recognition workshops , pages 1869–1871, 2018

    work page 2018

  8. [8]

    Momentum con- trast for unsupervised visual representation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum con- trast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729–9738, 2020

    work page 2020

  9. [9]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  10. [10]

    beta-vae: Learning basic visual concepts with a constrained variational framework

    Irina Higgins, Loic Matthey, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. In International conference on learning representations, 2017

    2017

  11. [11]

    Feature rep- resentation learning for unsupervised cross- domain image retrieval

    Conghui Hu and Gim Hee Lee. Feature rep- resentation learning for unsupervised cross- domain image retrieval. In European Con- ference on Computer Vision , pages 529–544. Springer, 2022

    work page 2022

  12. [12]

    Unsupervised feature representa- tion learning for domain-generalized cross- domain image retrieval

    Conghui Hu, Can Zhang, and Gim Hee Lee. Unsupervised feature representa- tion learning for domain-generalized cross- domain image retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 11016–11025, 2023

    work page 2023

  13. [13]

    Lora: Low- rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low- rank adaptation of large language models. ICLR, 1(2):3, 2022

    work page 2022

  14. [15]

    Cross-domain image retrieval with a dual attribute-aware ranking network

    Junshi Huang, Rogerio S Feris, Qiang Chen, and Shuicheng Yan. Cross-domain image retrieval with a dual attribute-aware ranking network. In Proceedings of the IEEE interna- tional conference on computer vision , pages 1062–1070, 2015

    work page 2015

  15. [16]

    Visual style prompting with swapping self-attention, 2024

    Jaeseok Jeong, Junho Kim, Yunjey Choi, Gayoung Lee, and Youngjung Uh. Visual style prompting with swapping self-attention, 2024

    work page 2024

  16. [17]

    Cross-domain image retrieval with attention modeling

    Xin Ji, Wei Wang, Meihui Zhang, and Yang Yang. Cross-domain image retrieval with attention modeling. In Proceedings of the 25th ACM international conference on Mul- timedia, pages 1654–1662, 2017

    work page 2017

  17. [18]

    Generative attribute con- troller with conditional filtered generative adversarial networks

    Takuhiro Kaneko, Kaoru Hiramatsu, and Kunio Kashino. Generative attribute con- troller with conditional filtered generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6089–6098, 2017

    work page 2017

  18. [19]

    Analyzing and improving the image quality of stylegan

    Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 8110–8119, 2020

    work page 2020

  19. [20]

    Ctrl: A conditional transformer language model for controllable generation

    Nitish Shirish Keskar, Bryan McCann, Lav R Varshney, Caiming Xiong, and Richard Socher. Ctrl: A conditional transformer language model for controllable generation. arXiv preprint arXiv:1909.05858 , 2019

    Pith/arXiv arXiv 1909

  20. [21]

    Cds: Cross-domain self- supervised pre-training

    Donghyun Kim, Kuniaki Saito, Tae-Hyun Oh, Bryan A Plummer, Stan Sclaroff, and Kate Saenko. Cds: Cross-domain self- supervised pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 9123–9132, 2021

    work page 2021

  21. [22]

    Picture that sketch: Photorealistic image generation from abstract sketches

    Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Picture that sketch: Photorealistic image generation from abstract sketches. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 6850–6861, June 2023

    work page 2023

  22. [23]

    A detect- then-retrieve model for multi-domain fash- ion item retrieval

    Michal Kucer and Naila Murray. A detect- then-retrieve model for multi-domain fash- ion item retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 14

    work page 2019

  23. [24]

    StyleID: Identity Disentanglement for Anonymizing Faces

    Minh-Ha Le and Niklas Carlsson. Styleid: Identity disentanglement for anonymizing faces. arXiv preprint arXiv:2212.13791, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  24. [25]

    Unsupervised cross-domain image retrieval via prototypical optimal transport

    Bin Li, Ye Shi, Qian Yu, and Jingya Wang. Unsupervised cross-domain image retrieval via prototypical optimal transport. In Pro- ceedings of the AAAI Conference on Artifi- cial Intelligence, volume 38, pages 3009–3017, 2024

    work page 2024

  25. [26]

    Prototypical contrastive learning of unsupervised representations

    Junnan Li, Pan Zhou, Caiming Xiong, and Steven CH Hoi. Prototypical contrastive learning of unsupervised representations. arXiv preprint arXiv:2005.04966 , 2020

    Pith/arXiv arXiv 2005

  26. [27]

    Dreamfont3d: personalized text-to-3d artistic font generation

    Xiang Li, Lei Meng, Lei Wu, Manyi Li, and Xiangxu Meng. Dreamfont3d: personalized text-to-3d artistic font generation. In ACM SIGGRAPH 2024 Conference Papers , pages 1–11, 2024

    work page 2024

  27. [28]

    Reducing domain gap in fre- quency and spatial domain for cross-modality domain adaptation on medical image segmen- tation

    Shaolei Liu, Siqi Yin, Linhao Qu, and Man- ning Wang. Reducing domain gap in fre- quency and spatial domain for cross-modality domain adaptation on medical image segmen- tation. In Proceedings of the AAAI Con- ference on Artificial Intelligence , volume 37, pages 1719–1727, 2023

    work page 2023

  28. [29]

    Deepfashion: Pow- ering robust clothes recognition and retrieval with rich annotations

    Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaoou Tang. Deepfashion: Pow- ering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1096–1104, 2016

    work page 2016

  29. [30]

    Specialist diffusion: Plug-and-play sample-efficient fine- tuning of text-to-image diffusion models to learn any unseen style

    Haoming Lu, Hazarapet Tunanyan, Kai Wang, Shant Navasardyan, Zhangyang Wang, and Humphrey Shi. Specialist diffusion: Plug-and-play sample-efficient fine- tuning of text-to-image diffusion models to learn any unseen style. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14267–14276, 2023

    work page 2023

  30. [31]

    Disentangling factors of variation in deep representation using adver- sarial training

    Michael F Mathieu, Junbo Jake Zhao, Junbo Zhao, Aditya Ramesh, Pablo Sprechmann, and Yann LeCun. Disentangling factors of variation in deep representation using adver- sarial training. Advances in neural informa- tion processing systems, 29, 2016

    work page 2016

  31. [32]

    Reducing domain gap by reducing style bias

    Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing domain gap by reducing style bias. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 8690–8699, 2021

    work page 2021

  32. [33]

    Representation learning with con- trastive predictive coding

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with con- trastive predictive coding. arXiv preprint arXiv:1807.03748, 2018

    Pith/arXiv arXiv 2018

  33. [34]

    Forgetmenot: Memory-aware forensic facial sketch match- ing

    Shuxin Ouyang, Timothy M Hospedales, Yi- Zhe Song, and Xueming Li. Forgetmenot: Memory-aware forensic facial sketch match- ing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recogni- tion, pages 5571–5579, 2016

    work page 2016

  34. [35]

    Moment matching for multi-source domain adapta- tion

    Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adapta- tion. In Proceedings of the IEEE Interna- tional Conference on Computer Vision, pages 1406–1415, 2019

    work page 2019

  35. [36]

    Test-time adaptation of discriminative models via diffusion genera- tive feedback

    Mihir Prabhudesai, Tsung-Wei Ke, Alexan- der Cong Li, Deepak Pathak, and Kate- rina Fragkiadaki. Test-time adaptation of discriminative models via diffusion genera- tive feedback. In Thirty-seventh Conference on Neural Information Processing Systems , 2023

    work page 2023

  36. [37]

    Ambiguous medical image segmenta- tion using diffusion models

    Aimon Rahman, Jeya Maria Jose Vala- narasu, Ilker Hacihaliloglu, and Vishal M Patel. Ambiguous medical image segmenta- tion using diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 11536– 11546, 2023

    work page 2023

  37. [38]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨ orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , pages 10684–10695, 2022

    work page 2022

  38. [39]

    Clip for all things zero-shot sketch-based image retrieval, fine-grained or not

    Aneeshan Sain, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. Clip for all things zero-shot sketch-based image retrieval, fine-grained or not. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 2765–2775, 2023

    work page 2023

  39. [40]

    Con- trolvae: Controllable variational autoencoder

    Huajie Shao, Shuochao Yao, Dachun Sun, Aston Zhang, Shengzhong Liu, Dongxin Liu, Jun Wang, and Tarek Abdelzaher. Con- trolvae: Controllable variational autoencoder. In International conference on machine 15 learning, pages 8655–8664. PMLR, 2020

    work page 2020

  40. [41]

    Content-based image retrieval at the end of the early years

    Arnold WM Smeulders, Marcel Worring, Simone Santini, Amarnath Gupta, and Ramesh Jain. Content-based image retrieval at the end of the early years. IEEE Trans- actions on pattern analysis and machine intelligence, 22(12):1349–1380, 2000

    work page 2000

  41. [42]

    Deep hashing network for unsupervised domain adaptation

    Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Pan- chanathan. Deep hashing network for unsupervised domain adaptation. In Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition , pages 5018–5027, 2017

    work page 2017

  42. [43]

    Idea: An invariant perspective for efficient domain adaptive image retrieval

    Haixin Wang, Hao Wu, Jinan Sun, Shikun Zhang, Chong Chen, Xian-Sheng Hua, and Xiao Luo. Idea: An invariant perspective for efficient domain adaptive image retrieval. Advances in Neural Information Processing Systems, 36:57256–57275, 2023

    work page 2023

  43. [45]

    Correspondence-free domain align- ment for unsupervised cross-domain image retrieval

    Xu Wang, Dezhong Peng, Ming Yan, and Peng Hu. Correspondence-free domain align- ment for unsupervised cross-domain image retrieval. In Proceedings of the AAAI Con- ference on Artificial Intelligence , volume 37, pages 10200–10208, 2023

    work page 2023

  44. [46]

    Styleadapter: A unified styl- ized image generation model

    Zhouxia Wang, Xintao Wang, Liangbin Xie, Zhongang Qi, Ying Shan, Wenping Wang, and Ping Luo. Styleadapter: A unified styl- ized image generation model. arXiv preprint arXiv:2309.01770, 2023

    Pith/arXiv arXiv 2023

  45. [47]

    Tune-a-video: One- shot tuning of image diffusion models for text-to-video generation

    Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One- shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 7623–7633, 2023

    work page 2023

  46. [48]

    Unsupervised feature learn- ing via non-parametric instance discrimina- tion

    Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learn- ing via non-parametric instance discrimina- tion. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 3733–3742, 2018

    work page 2018

  47. [49]

    Self-supervised domain adaptation for computer vision tasks

    Jiaolong Xu, Liang Xiao, and Antonio M L´ opez. Self-supervised domain adaptation for computer vision tasks. IEEE Access , 7:156694–156706, 2019

    work page 2019

  48. [50]

    A fourier-based framework for domain generalization

    Qinwei Xu, Ruipeng Zhang, Ya Zhang, Yan- feng Wang, and Qi Tian. A fourier-based framework for domain generalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 14383–14392, 2021

    work page 2021

  49. [51]

    Learning dynamic style kernels for artistic style transfer

    Wenju Xu, Chengjiang Long, and Yongwei Nie. Learning dynamic style kernels for artistic style transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 10083–10092, 2023

    work page 2023

  50. [52]

    Representation align- ment for generation: Training diffusion trans- formers is easier than you think

    Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation align- ment for generation: Training diffusion trans- formers is easier than you think. arXiv preprint arXiv:2410.06940, 2024

    Pith/arXiv arXiv 2024

  51. [53]

    Prototypical cross-domain self- supervised learning for few-shot unsupervised domain adaptation

    Xiangyu Yue, Zangwei Zheng, Shanghang Zhang, Yang Gao, Trevor Darrell, Kurt Keutzer, and Alberto Sangiovanni Vin- centelli. Prototypical cross-domain self- supervised learning for few-shot unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 13834–13844, 2021

    work page 2021

  52. [54]

    A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence

    Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. Advances in Neural Information Processing Systems, 36:45533–45547, 2023

    2023

  53. [55]

    Controlvideo: Training-free control- lable text-to-video generation

    Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free control- lable text-to-video generation. arXiv preprint arXiv:2305.13077, 2023

    Pith/arXiv arXiv 2023

  54. [56]

    Cross-domain image retrieval: methods and applications

    Xiaoping Zhou, Xiangyu Han, Haoran Li, Jia Wang, and Xun Liang. Cross-domain image retrieval: methods and applications. Inter- national Journal of Multimedia Information Retrieval, 11(3):199–218, 2022. 16

    work page 2022

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.