Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Generating a Paracosm for Training-Free Zero-Shot Composed Image Retrieval

T0 review · 3 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read A training-free method that renders the query's 'mental image' and synthetic database twins reaches state-of-the-art zero-shot composed image retrieval.

desk verdict Training-free CIR by generating a mental image plus synthetic database counterparts; the SOTA results look credible but lack variance reports and a contamination check. read the letter →

arxiv 2602.00813 v6 pith:P2ACHJDU submitted 2026-01-31 cs.CV

Tong Wang , Yunhan Zhao , Shu Kong This is my paper
classification cs.CV
keywords composedimageretrievalzero-shottraining-freemethodmentalgenerationsynthetic-to-realdomaingaplargemultimodalmodelseditingCLIPfeaturematching
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

Paracosm tries to settle the core problem of composed image retrieval (CIR) — finding a photo from a reference image plus text saying how to change it — the direct way: instead of describing the intended result in words, it makes an image generator produce the 'mental image' the query implies. Because that rendering is synthetic while the database is real, it also renders a synthetic counterpart of every database image, so query and database live in one shared synthetic world ('paracosm') where a frozen CLIP encoder can match them image-to-image. The paper reports that this training-free recipe beats every existing zero-shot CIR method on CIRR, CIRCO, and FashionIQ, including training-based ones, and rivals supervised methods. Why a sympathetic reader would care: it suggests triplet-supervised training, long treated as necessary for CIR, can be bypassed entirely by generative models, and that retrieval quality now tracks image-editing fidelity.

What carries the argument

The load-bearing object is the paracosm: a matched pair of synthetic images that makes retrieval symmetric. On the query side, an image-editing large multimodal model rewrites the reference image according to the modification text, producing a 'mental image', and a second prompt condenses that mental image into a one-sentence description. On the database side, a captioning model writes a detailed description of each real photo and a text-to-image model re-renders it as a synthetic counterpart. The query feature is the mental image's visual feature plus its description's text feature plus the modification text, and each database feature is the real photo plus the synthetic twin; retrieval is

What would settle it

Measure the CLIP cosine distance between each real database image and its synthetic counterpart: if a large share of targets are farther from their own twins than from unrelated images, the paracosm bridge is not actually carrying identity information. The paper's Fig. 7 already supplies the companion signal — retrieval fails precisely where the editor fabricates a detail or skips an edit — so a structured test set separating 'edit succeeded' from 'edit failed' queries (e.g., exact counts and role swaps) should reproduce that collapse quantitatively.

Watch

Extended reading notes

Core claim

Paracosm's central claim is that the bottleneck in zero-shot CIR is not how to fuse reference and text features, but that the target's visual form is missing: text-only descriptions of the query discard the fine-grained cues in the reference photo. The method constructs what it calls a paracosm — a synthetic universe in which the multimodal query is rendered as an edited 'mental image' (an image-editing large multimodal model, Qwen-Image-Edit in the paper, applies the modification text to the reference photo), and every database image is re-rendered as a synthetic counterpart generated from its detailed caption. Matching then reduces to cosine similarity between a query feature (mental image

Load-bearing premise

The pipeline stands on one premise: the image generators preserve the exact attributes the modification text cares about — counts, colors, positions, swapped roles — closely enough that CLIP features of the synthetic mental image and the synthetic database twins land near their real counterparts; the paper's own failure figure shows this premise breaking when the editor invents a cartoon duck, appends counterfactual gas burners, or ignores a requested color change.

Editorial extensions

If this is right

  • If Paracosm is right, zero-shot CIR no longer needs indirect training pipelines: an off-the-shelf editing model, a captioner, a text-to-image model, and a frozen CLIP encoder replace textual-inversion training entirely.
  • Directly editing the reference image beats generating a pseudo-target from a rewritten description (Table 2), so query rendering should start from the actual reference photo rather than from text.
  • Synthetic counterparts of database images are worth generating even at the cost of roughly 13 GPU-hours over a 123K-image database, since they can be precomputed once and stored as compact features.
  • Retrieval failures concentrate where the generator fabricates or misses details (Fig. 7), implying that benchmark scores of CIR methods now upper-bound the fidelity of the underlying image editor.
  • Because a single weighting choice works across all three benchmarks without per-dataset tuning, the method transfers across domains with minimal hyperparameter adjustment.

Reading between the lines

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

  • A natural extension the paper leaves implicit: retrieve by edit-consistency — re-edit the top candidates back toward the reference image and rank by how well the round trip reproduces it — which would neutralize exactly the counterfactual details (the invented duck, the extra gas burners) that Fig. 7 shows derailing retrieval.
  • If the paracosm mechanism is the real driver, the same recipe should transfer to other asymmetric-domain retrieval tasks, such as sketch-to-photo or rendered-to-real product search, where one side of the comparison is inherently synthetic.
  • The method makes a falsifiable scaling prediction: retrieval gains should track how close each synthetic database twin is to its real photo in CLIP space, a cheap per-image distance that could serve as a quality gate before retrieval.
  • Because the query feature sums the mental image and its description, failures might be diagnosable by scoring each component separately per query to see which one pulls the query toward a wrong hit.
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

3 major / 4 minor

Summary. The paper proposes Paracosm, a training-free zero-shot composed image retrieval (ZS-CIR) method. For a query consisting of a reference image and modification text, an LMM-based image editor (Qwen-Image-Edit) generates a 'mental image'; Qwen2.5-VL then produces a one-sentence description of that mental image. For each database image, Qwen2.5-VL generates a detailed caption, and Qwen-Image generates a synthetic counterpart image. Retrieval uses CLIP features combined as: query feature q = λ(V(I_mental)+T(t_query))+(1−λ)T(t_mod), database feature φ_i = V(I_i)+V(I_i^syn), followed by cosine similarity. The hyperparameter λ=0.3 is tuned on the CIRR validation set, and β=0.5 is used in the supplementary for a weighted counterpart variant. Experiments on CIRR, CIRCO, and FashionIQ, with ViT-L/14 and ViT-G/14 backbones, report state-of-the-art zero-shot results and even competitiveness with some supervised methods. Ablations in Table 5 show that removing the mental image, synthetic counterparts, or modification text degrades performance.

Significance. If the reported results hold, the paper makes a significant empirical contribution: it demonstrates that a purely training-free, generative pipeline can outperform training-based zero-shot CIR methods and rival supervised approaches. The central idea of constructing a synthetic 'paracosm'---matching synthetic query images to synthetic database counterparts---is well motivated and goes beyond the text-only descriptions used by prior training-free methods. The paper is clearly written, the ablations are internally consistent, and the authors provide code and a demo notebook. However, the headline SOTA claim rests on empirical evidence that currently lacks two pieces of support: a test against data contamination in the generative models, and a statistical treatment of stochastic generation. Both are addressable within the manuscript's scope.

major comments (3)
  1. [§3.2, Eq. (1), Tables 3-4] The central SOTA claim assumes that Qwen-Image/Qwen-Image-Edit generate faithful synthetic images rather than near-duplicates of benchmark images. CIRR, CIRCO, and FashionIQ are public and likely appear in web-scale pretraining data. Since Eq. (1) adds V(I_i^syn) to V(I_i) and V(I_mental) to the query, memorization would trivially inflate retrieval scores, e.g., CIRCO mAP@5 39.82 vs. 32.23 for CoTMR. The paper provides no membership test, no similarity analysis between generated and real images, and no evaluation on a database outside the generators' training distribution. Please add nearest-neighbor similarity distributions, retrieval-level leakage tests, or results on a private/held-out split before accepting the SOTA claim.
  2. [§4, Tables 3-4] All reported numbers are single-run, despite the image-generation pipeline being stochastic (Qwen-Image-Edit and Qwen-Image use nondeterministic sampling). Some gaps are small, e.g., FashionIQ Dress R@10 33.17 vs. 33.02 for OSrCIR (Table 4) and CIRR Rsubset@1 70.82 vs. 69.22 for OSrCIR (Table 3). The phrase 'significantly outperforms' is therefore not statistically supported. Please report mean ± std over multiple generation seeds and database-processing runs, ideally with a paired significance test over the query set.
  3. [§3.2, Fig. 4, Limitations] Paracosm uses dataset-specific prompt templates, as acknowledged in the Limitations. Some adaptation to the format of modification texts is reasonable, but it makes the comparison against methods that use a generic prompt not fully controlled. To make the SOTA claim attributable to the method rather than to per-dataset prompt engineering, please include an ablation with a single fixed prompt across all datasets, or quantify the gain from template adaptation.
minor comments (4)
  1. [Supplementary Sec. C vs. Eq. (1)] The text says the weighting parameter β in 'Eq. (1)' is set to 0.5, but the main-paper Eq. (1) contains no β; the weighted variant should be written out explicitly in the main text or the reference should be corrected.
  2. [Fig. 1 caption] Typo: 'synesthetic counterparts' should read 'synthetic counterparts'.
  3. [Fig. 6] Axis label 'Retrival Performance' should be 'Retrieval Performance'.
  4. [Table 5] The checkmark alignment between row entries and column headers is ambiguous in the first block. Please restructure the table or key each row to the exact terms in Eq. (1), e.g., with a compact notation for active components.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: Paracosm's SOTA claim rests on external benchmarks; lambda is ordinary hyperparameter tuning and synthetic counterparts are input-derived feature augmentation, not fitted predictions.

full rationale

The paper's derivation chain is not circular. The central claim is an empirical result on held-out test sets (CIRR, CIRCO, FashionIQ) using fixed pretrained CLIP/OpenCLIP encoders. The only tuned hyperparameter is λ=0.3, selected on the CIRR validation set and then fixed; this is standard hyperparameter selection, not a prediction forced by construction, and the paper reports gains on test sets outside the tuning split. The synthetic counterparts I_syn are generated from each database image I_i and added to its feature in Eq. (1) (ϕ_i = V(I_i) + V(I_syn)); this is feature augmentation derived from the database item itself, not from target labels or retrieval ground truth, and Table 5 ablates its contribution, showing an independent measurable effect. The 'mental image' is generated from the reference image plus modification text without access to the target. Self-citations (e.g., refs. [22], [45], [67]) appear as related-work motivation for synthetic-to-real alignment, but the load-bearing evidence is the paper's own ablations and external benchmarks, not a self-citation chain. The skeptical concern about possible memorization/contamination of benchmark images by the Qwen generators is a question of external validity and data leakage, not of the derivation reducing to its inputs; it would require a membership test to substantiate and does not constitute circularity under the stated rules. The paper's own stated limitations (Sec. 5, Fig. 7) acknowledge generative infidelity and retrieval failures, which further confirms that the method's outputs are not definitionally guaranteed to match the targets.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

No new physical or theoretical entities are introduced; "mental image" and "synthetic counterpart" are generated artifacts rather than new scientific constructs. The method's load-bearing inputs are all pretrained foundation models, manual prompt templates, and two weighting hyperparameters.

free parameters (2)
  • lambda = 0.3
    Eq. (1) weight for combining mental-image/query-description features vs modification text; tuned on the CIRR validation set with CLIP ViT-B/32 (Sec. 4.2, Fig. 6).
  • beta = 0.5
    Supplementary Fig. 9 balance weight between real database image features and synthetic counterpart features; set to 0.5 based on CIRR/CIRCO validation experiments and for model simplicity.
assumptions (5)
  • domain assumption Qwen-Image-Edit can faithfully apply the modification text to the reference image to produce a mental image preserving identity-bearing details.
    Central to the query representation; the paper's own Fig. 7 shows failures where the model adds a cartoon duck, invents counterfactual gas burners, or ignores a color edit.
  • domain assumption Qwen-Image synthetic counterparts preserve enough fine-grained discriminative information of real database images to match mental images in CLIP space.
    Eq. (1) adds V(I_syn) to database features; if synthetic counterparts lose fine-grained attributes, retrieval quality degrades.
  • domain assumption CLIP embeddings of real and synthetic images are comparable enough for cosine retrieval in Eq. (1) to be meaningful across the synthetic-to-real gap.
    The entire paracosm matching strategy relies on pretrained VLM features bridging both domains; this is asserted, not derived.
  • domain assumption Pretrained LMMs and T2I generators (Qwen2.5-VL, Qwen-Image, Qwen-Image-Edit) were not trained on or heavily memorized CIRR/CIRCO/FashionIQ evaluation images.
    Zero-shot claims would be inflated by contamination; the paper provides no contamination analysis in Sec. 4.
  • domain assumption Detailed descriptions generated by Qwen2.5-VL capture all visible objects, attributes, and spatial relationships needed for high-fidelity synthetic counterparts.
    Database preprocessing in Fig. 5 depends on description completeness; imperfect descriptions will propagate into synthetic counterparts.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generating a Paracosm for Training-Free Zero-Shot Composed Image Retrieval." pith.science (2026). https://pith.science/paper/P2ACHJDU

@misc{pith2026260200813,
  author       = {Pith},
  title        = {Pith review of: Generating a Paracosm for Training-Free Zero-Shot Composed Image Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P2ACHJDU}},
  note         = {Machine review of arXiv:2602.00813}
}
read the original abstract

Composed Image Retrieval (CIR) is the task of retrieving a target image from a database using a multimodal query, which consists of a reference image and a modification text. The text specifies how to alter the reference image to form a ''mental image'', based on which CIR should find the target image in the database. The fundamental challenge of CIR is that this ''mental image'' is not physically available and is only implicitly defined by the query. The contemporary literature pursues zero-shot methods and uses a Large Multimodal Model (LMM) to generate a textual description for a given multimodal query, and then employs a Vision-Language Model (VLM) for textual-visual matching to search for the target image. In contrast, we address CIR from first principles by directly generating the ''mental image'' for more accurate matching. Particularly, we prompt an LMM to generate a ''mental image'' for a given multimodal query and propose to use this ''mental image'' to search for the target image. As the ''mental image'' has a synthetic-to-real domain gap with real images, we also generate a synthetic counterpart for each real image in the database to facilitate matching. In this sense, our method uses LMM to construct a ``paracosm'', where it matches the multimodal query and database images. Hence, we call this method Paracosm. Notably, Paracosm is a training-free zero-shot CIR method. It significantly outperforms existing zero-shot methods on challenging benchmarks, achieving state-of-the-art performance for zero-shot CIR.

Figures

Figures reproduced from arXiv: 2602.00813 by the authors.

Figure 1
Figure 1. Overview of our method and benchmarking results. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Flowchart of our training-free zero-shot CIR method Paracosm. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Comparison of qualitative results between OSrCIR [49] and our Paracosm. We show four examples from the CIRCO dataset [1] in the first column, followed by generated descriptions and top-4 retrievals by OSrCIR, and the mental images and top-4 retrievals by Paracosm. For each multimodal query, OSrCIR uses an LMM to generate a description, uses it to match database images, and returns top-ranked ones. Instead, Paracosm … view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Illustration of processing a multimodal query. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Illustration of processing database images. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Analysis of λ which controls the importance of incorporating modification text in Eq. 1. We set λ = 0.3 based on the results on the CIRR validation set. Interestingly, on all datasets, setting λ = 0.3 consistently yields the highest numeric metrics reported for all the…
Figure 7
Figure 7. Figure 7: Failure cases on the CIRCO dataset. Paracosm can fail due to limitations of generative models that can generate implausible and counterfactual mental images. Four examples, respectively, demonstrate different failures of Paracosm. (1) It incorrectly generates a cartoon…
Figure 8
Figure 8. Figure 8: Extended flowchart of Paracosm incorporating brief database image [PITH_FULL_IMAGE:figures/full_fig_p022_8.png]
Figure 10
Figure 10. Figure 10: Additional qualitative examples of Paracosm [PITH_FULL_IMAGE:figures/full_fig_p025_10.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Controlling Embedding Spaces with Text-Conditioned Transformations

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A single hypernetwork turns text descriptions of attributes into affine maps of frozen CLIP embeddings, making those attributes control retrieval and clustering without re-encoding the gallery.

Reference graph

Works this paper leans on

68 extracted references · 10 linked inside Pith · cited by 1 Pith paper

  1. [1]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)47(11), 10801–10817 (2025).https: //doi.org/10.1109/TPAMI.2025.3593539

    Agnolucci, L., Baldrati, A., Bimbo, A.D., Bertini, M.: isearle: Improving textual inversion for zero-shot composed image retrieval. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)47(11), 10801–10817 (2025).https: //doi.org/10.1109/TPAMI.2025.3593539

    arXiv 2025

  2. [2]

    arXiv preprint arXiv:2309.16609 (2023)

    Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., et al.: Qwen technical report. arXiv preprint arXiv:2309.16609 (2023)

    arXiv 2023

  3. [3]

    5-vl technical report

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., et al.: Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923 (2025)

    arXiv 2025

  4. [4]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Baldrati, A., Agnolucci, L., Bertini, M., Del Bimbo, A.: Zero-shot composed image retrieval with textual inversion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15338–15347 (October 2023)

    2023

  5. [5]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Baldrati, A., Bertini, M., Uricchio, T., Del Bimbo, A.: Effective conditioned and composed image retrieval combining clip-based features. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21466–21474 (June 2022)

    2022

  6. [6]

    In: Proceedings of the 31st International Conference on Computational Linguistics (COLING)

    Bao, T., Liu, C., Xu, D., Zheng, Z., Xu, T.: Mllm-i2w: Harnessing multimodal large language model for zero-shot composed image retrieval. In: Proceedings of the 31st International Conference on Computational Linguistics (COLING). pp. 1839–1849 (2025)

    2025

  7. [7]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18392–18402 (June 2023)

    2023

  8. [8]

    Advances in Neural Information Processing Systems (NeurIPS)33, 1877– 1901 (2020)

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Nee- lakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS)33, 1877– 1901 (2020)

    1901

Show all 68 references
  1. [9]

    In: European Conference on Computer Vision (ECCV)

    Chen, Y., Bazzani, L.: Learning joint visual semantic matching embeddings for language-guided retrieval. In: European Conference on Computer Vision (ECCV). pp. 136–152. Springer (2020)

    2020

  2. [10]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Chen, Y., Gong, S., Bazzani, L.: Image search with text feedback by visiolinguistic attention learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3001–3011 (2020)

    2020

  3. [11]

    In: International Conference on Learning Representations (ICLR) (2024)

    Chen, Y., Zheng, Z., Ji, W., Qu, L., Chua, T.S.: Composed image retrieval with text feedback via multi-grained uncertainty regularization. In: International Conference on Learning Representations (ICLR) (2024)

    2024

  4. [12]

    In: Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD)

    Cheng, Z., Ma, Y., Lang, J., Zhang, K., Zhong, T., Wang, Y., Zhou, F.: Generative thinking, corrective action: User-friendly composed image retrieval via automatic multi-agent collaboration. In: Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mini...

    2025

  5. [13]

    In: International Confer- ence on Learning Representations (ICLR) (2022)

    Delmas, G., Rezende, R.S., Csurka, G., Larlus, D.: Artemis: Attention-based re- trieval with text-explicit matching and implicit similarity. In: International Confer- ence on Learning Representations (ICLR) (2022)

    2022

  6. [14]

    https://github.com/modelscope/ diffsynth-engine(2025)

    Di, Z., Zhu, G., Duan, Z., Chu, Z., Chen, Y., Lu, W.: Diffsynth-engine: a high-performance diffusion inference engine. https://github.com/modelscope/ diffsynth-engine(2025)

    2025

  7. [15]

    In: Proceedings of the 16 T

    Gao, J., Zhang, J., Liu, X., Darrell, T., Shelhamer, E., Wang, D.: Back to the source: Diffusion-driven adaptation to test-time corruption. In: Proceedings of the 16 T. Wang et al. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)

    2023

  8. [16]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Gu, G., Chun, S., Kim, W., Kang, Y., Yun, S.: Language-only training of zero- shot composed image retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13225–13234 (June 2024)

    2024

  9. [17]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR) (2025)

    Guo, J., Zhao, J., Du, C., Wang, Y., Ge, C., Ni, Z., Song, S., Shi, H., Huang, G.: Everything to the synthetic: Diffusion-driven test-time adaptation via synthetic- domain alignment. In: Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR) (2025)

    2025

  10. [18]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Hu, Z., Zhu, X., Tran, S., Vidal, R., Dhua, A.: Provla: Compositional image search with progressive vision-language alignment and multimodal fusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2772– 2777 (2023)

    2023

  11. [19]

    Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021).https://doi.org/10.5281/zenodo.5143773 , https://doi.org/10.5281/zenodo.5143773

    2021 doi

  12. [20]

    arXiv preprint arXiv:2408.07009 (2024)

    Imagen-Team-Google: Imagen 3. arXiv preprint arXiv:2408.07009 (2024)

    2024

  13. [21]

    In: International Conference on Learning Represen- tations (ICLR) (2024)

    Karthik, S., Roth, K., Mancini, M., Akata, Z.: Vision-by-language for training-free compositional image retrieval. In: International Conference on Learning Represen- tations (ICLR) (2024)

    2024

  14. [22]

    In: Proceedings of the CHI Conference on Human Factors in Computing Systems (2024)

    Kwon, N., Lu, Q., Qazi, M.H., Liu, J., Oh, C., Kong, S., Kim, J.: Accesslens: Auto- detecting inaccessibility of everyday objects. In: Proceedings of the CHI Conference on Human Factors in Computing Systems (2024)

    2024

  15. [23]

    Labs, B.F.: FLUX.2: Frontier Visual Intelligence.https://bfl.ai/blog/flux-2 (2025)

    2025

  16. [24]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Lee, S., Kim, D., Han, B.: Cosmo: Content-style modulation for image retrieval with text feedback. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 802–812 (June 2021)

    2021

  17. [25]

    In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI)

    Levy, M., Ben-Ari, R., Darshan, N., Lischinski, D.: Data roaming and quality assessment for composed image retrieval. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 38, pp. 2991–2999 (2024)

    2024

  18. [26]

    In: International Conference on Machine Learning (ICML)

    Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning (ICML). pp. 12888–12900. PMLR (2022)

    2022

  19. [27]

    In: Proceedings of the 2024 International Conference on Multimedia Retrieval (ICMR)

    Li, M., Zhao, Z., Jiang, X., Jiang, Z.: Clip-probcr: Clip-based probability embedding combination retrieval. In: Proceedings of the 2024 International Conference on Multimedia Retrieval (ICMR). pp. 1104–1109 (2024)

    2024

  20. [28]

    In: Forty-first International Conference on Machine Learning (ICML) (2024)

    Li, W., Fan, H., Wong, Y., Yang, Y., Kankanhalli, M.: Improving context under- standing in multimodal large language models via multimodal composition learning. In: Forty-first International Conference on Machine Learning (ICML) (2024)

    2024

  21. [29]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Li, Y., Ma, F., Yang, Y.: Imagine and seek: Improving composed image retrieval with an imagined proxy. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3984–3993 (June 2025)

    2025

  22. [30]

    In: European Conference on Computer Vision (ECCV)

    Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (ECCV). pp. 740–755. Springer (2014)

    2014

  23. [31]

    Liu, Y., Yao, J., Zhang, Y., Wang, Y., Xie, W.: Zero-shot composed text-image retrieval (2024),https://arxiv.org/abs/2306.07272 Paracosm 17

    2024 arXiv

  24. [32]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Liu,Z.,Rodriguez-Opazo,C.,Teney,D.,Gould,S.:Imageretrievalonreal-lifeimages with pre-trained vision-and-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2125–2134 (October 2021)

    2021

  25. [33]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)

    Liu, Z., Sun, W., Hong, Y., Teney, D., Gould, S.: Bi-directional training for composed image retrieval via text prompt learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 5753–5762 (January 2024)

    2024

  26. [34]

    Transactions on Machine Learning Research (TMLR) (2024),https://openreview.net/forum?id=fJAwemcvpL

    Liu, Z., Sun, W., Teney, D., Gould, S.: Candidate set re-ranking for composed image retrieval with dual multi-modal encoder. Transactions on Machine Learning Research (TMLR) (2024),https://openreview.net/forum?id=fJAwemcvpL

    2024

  27. [35]

    OpenAI: Gpt-4 technical report (2024),https://arxiv.org/abs/2303.08774

    2024 arXiv

  28. [36]

    In: International Conference on Machine Learning (ICML)

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (ICML). pp. 8748–8763. PmLR (2021)

    2021

  29. [37]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684–10695 (June 2022)

    2022

  30. [38]

    International Journal of Computer Vision (IJCV)115(3), 211–252 (2015).https://doi.org/10.1007/s11263-015-0816-y

    Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV)115(3), 211–252 (2015).https://do...

    2015 doi

  31. [39]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Saito, K., Sohn, K., Zhang, X., Li, C.L., Lee, C.Y., Saenko, K., Pfister, T.: Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19305–19314 (June 2023)

    2023

  32. [40]

    In: European Conference on Computer Vision (ECCV)

    Sauer, A., Lorenz, D., Blattmann, A., Rombach, R.: Adversarial diffusion distillation. In: European Conference on Computer Vision (ECCV). pp. 87–103. Springer (2024)

    2024

  33. [41]

    In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (ACL)

    Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (ACL). pp. 2556–2565 (2018)

    2018

  34. [42]

    In: The 41st International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR)

    Song, X., Feng, F., Han, X., Yang, X., Liu, W., Nie, L.: Neural compatibility modeling with attentive knowledge distillation. In: The 41st International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR). pp. 5–14 (2018)

    2018

  35. [43]

    In: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL)

    Suhr, A., Zhou, S., Zhang, A., Zhang, I., Bai, H., Artzi, Y.: A corpus for reasoning about natural language grounded in photographs. In: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL). pp. 6418–6428 (2019)

    2019

  36. [44]

    In: Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision (ICCV)

    Sun, Z., Jing, D., Lu, Z.: Cotmr: Chain-of-thought multi-scale reasoning for training- free zero-shot composed image retrieval. In: Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision (ICCV). pp. 22675–22684 (October 2025)

    2025

  37. [45]

    In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition (CVPR) (2024) 18 T

    Sun, Z., Fang, Y., Wu, T., Zhang, P., Zang, Y., Kong, S., Xiong, Y., Lin, D., Wang, J.: Alpha-clip: A clip model focusing on wherever you want. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition (CVPR) (2024) 18 T. Wang et al

    2024

  38. [46]

    arXiv preprint arXiv:1908.07490 (2019)

    Tan, H., Bansal, M.: Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490 (2019)

    1908 arXiv

  39. [47]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)

    Tang, Y., Yu, J., Gai, K., Zhuang, J., Xiong, G., Gou, G., Wu, Q.: Missing target- relevant information prediction with world model for accurate zero-shot composed image retrieval. In: Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR). pp. 24785–24795 (2025)

    2025

  40. [48]

    In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI)

    Tang, Y., Yu, J., Gai, K., Zhuang, J., Xiong, G., Hu, Y., Wu, Q.: Context-i2w: Mapping images to context-dependent words for accurate zero-shot composed image retrieval. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 38, pp. 5180–5188 (2024)

    2024

  41. [49]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Tang, Y., Zhang, J., Qin, X., Yu, J., Gou, G., Xiong, G., Lin, Q., Rajmohan, S., Zhang, D., Wu, Q.: Reason-before-retrieve: One-stage reflective chain-of-thoughts for training-free zero-shot composed image retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Visio...

    2025

  42. [50]

    arXiv preprint arXiv:2312.11805 (2023)

    Team, G.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)

    2023 arXiv

  43. [51]

    arXiv preprint arXiv:2512.07584 (2025)

    Team, M.L., Ma, H., Tan, H., Huang, J., Wu, J., He, J.Y., Gao, L., Xiao, S., Wei, X., Ma, X., et al.: Longcat-image technical report. arXiv preprint arXiv:2512.07584 (2025)

    2025 arXiv

  44. [52]

    Team, Q.: Qwen2.5-vl (January 2025),https://qwenlm.github.io/blog/qwen2.5- vl/

    2025

  45. [53]

    In: Proceedings of the Winter Conference on Applications of Computer Vision (WACV)

    Tian, Y., Newsam, S., Boakye, K.: Fashion image retrieval with text feedback by additive attention compositional learning. In: Proceedings of the Winter Conference on Applications of Computer Vision (WACV). pp. 1011–1021 (2023)

    2023

  46. [54]

    arXiv preprint arXiv:2302.13971 (2023)

    Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

    2023 arXiv

  47. [55]

    arXiv preprint arXiv:2307.09288 (2023)

    Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)

    2023 arXiv

  48. [56]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)

    Vo, N., Jiang, L., Sun, C., Murphy, K., Li, L.J., Fei-Fei, L., Hays, J.: Composing text and image for image retrieval - an empirical odyssey. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)

    2019

  49. [57]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Wang, L., Ao, W., Boddeti, V.N., Lim, S.N.: Generative zero-shot composed image retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 29690–29700 (June 2025)

    2025

  50. [58]

    Wu, C., Li, J., Zhou, J., Lin, J., Gao, K., Yan, K., ming Yin, S., Bai, S., Xu, X., Chen, Y., Chen, Y., Tang, Z., Zhang, Z., Wang, Z., Yang, A., Yu, B., Cheng, C., Liu, D., Li, D., Zhang, H., Meng, H., Wei, H., Ni, J., Chen, K., Cao, K., Peng, L., Qu, L., Wu, M., Wang, P., Yu,...

    2025 arXiv

  51. [59]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Wu, H., Gao, Y., Guo, X., Al-Halah, Z., Rennie, S., Grauman, K., Feris, R.: Fashion iq: A new dataset towards retrieving images by natural language feedback. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11307–11317 (June 202...

    2021

  52. [60]

    IEEE Transactions on Multimedia (TMM)25, 8346–8357 (2023)

    Xu, Y., Bin, Y., Wei, J., Yang, Y., Wang, G., Shen, H.T.: Multi-modal transformer with global-local alignment for composed query image retrieval. IEEE Transactions on Multimedia (TMM)25, 8346–8357 (2023)

    2023

  53. [61]

    In: Findings of the Association for Computational Linguistics (ACL)

    Yang, Y., Wang, Y., Wang, Y.: Sda: Semantic discrepancy alignment for text- conditioned image retrieval. In: Findings of the Association for Computational Linguistics (ACL). pp. 5250–5261 (2024)

    2024

  54. [62]

    In: Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR)

    Yang, Z., Xue, D., Qian, S., Dong, W., Xu, C.: Ldre: Llm-based divergent reasoning and ensemble for zero-shot composed image retrieval. In: Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR). pp. 80–90 (2024)

    2024

  55. [63]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Yin, T., Gharbi, M., Zhang, R., Shechtman, E., Durand, F., Freeman, W.T., Park, T.: One-step diffusion with distribution matching distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6613–6623 (2024)

    2024

  56. [64]

    In: Salakhut- dinov, R., Kolter, Z., Heller, K., Weller, A., Oliver, N., Scarlett, J., Berkenkamp, F

    Zhang, K., Luan, Y., Hu, H., Lee, K., Qiao, S., Chen, W., Su, Y., Chang, M.W.: Mag- icLens: Self-supervised image retrieval with open-ended instructions. In: Salakhut- dinov, R., Kolter, Z., Heller, K., Weller, A., Oliver, N., Scarlett, J., Berkenkamp, F. (eds.) International ...

    2024

  57. [65]

    Knowledge-Based Systems (KBS)300, 112135 (2024)

    Zhang, X., Zheng, Z., Zhu, L., Yang, Y.: Collaborative group: Composed image retrieval via consensus learning from noisy annotations. Knowledge-Based Systems (KBS)300, 112135 (2024)

    2024

  58. [66]

    In: Proceedings of the 45th international ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR)

    Zhao, Y., Song, Y., Jin, Q.: Progressive learning for image retrieval with hybrid- modality queries. In: Proceedings of the 45th international ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR). pp. 1012–1021 (2022)

    2022

  59. [67]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)

    Zhao,Y.,Kong,S.,Shin,D.,Fowlkes,C.:Domaindecluttering:Simplifyingimagesto mitigate synthetic-real domain shift and improve depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)

    2020

  60. [68]

    paracosm

    Zhou, D., Li, Y., Ma, F., Zhang, X., Yang, Y.: Migc: Multi-instance generation controller for text-to-image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6818–6828 (June 2024) Paracosm 1 Generating a Paracosm for T...

    2024

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.