Pith. sign in

REVIEW 4 major objections 3 minor 56 references

FACap: A Large-scale Fashion Dataset for Fine-grained Composed Image Retrieval

T0 review · 4 major / 3 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 The paper introduces FACap, a 227,680-triplet fashion composed-image-retrieval dataset built automatically by a VLM and an LLM, and a BLIP-2-based model that uses it to beat zero-shot baselines and reach near-state-of-the-art fine-tuned…

desk verdict FACap is a real resource with clean ablations, but the headline fine-grained numbers come from a benchmark the authors built with the same pipeline that made the training set, so the 'especially fine-grained' claim is conditional. read the letter →

arxiv 2507.07135 v1 pith:F7EB7Y7Z submitted 2025-07-08 cs.LG

François Gardères , Shizhe Chen , Camille-Sovanneary Gauthier , Jean Ponce This is my paper
classification cs.LG
keywords ComposedimageretrievalMultimodalfusionFashiondomainDatasetconstructionVision-languagemodelsFine-grainedAutomaticannotationBLIP-2
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

The paper sets out to show that the bottleneck in fashion composed image retrieval—finding a target image from a reference image plus a short text describing changes—is data, not just architecture. It introduces FACap, a dataset of 227,680 automatically constructed triplets built by pairing web-sourced fashion images, captioning each image with a vision-language model, and having a large language model write the modification texts that describe the differences. It then introduces FashionBLIP-2, which adapts the pretrained BLIP-2 model with lightweight adapters and multi-head query-candidate matching instead of a single global vector. Reported results: FashionBLIP-2 pretrained only on FACap averages 44.63 Recall on FashionIQ, above the 43.97 of the best zero-shot method, and after FashionIQ fine-tuning reaches 65.97, close to the 67.93 of the strongest existing model. A sympathetic reading is that automatically generated, fine-grained annotations can substitute for costly specialist labeling in fashion retrieval at scale.

What carries the argument

The load-bearing mechanism is the two-stage annotation pipeline. In stage one, the VLM InternVL captions each image in up to 128 tokens, prompted with the image category and noisy product metadata to reduce hallucination. In stage two, the LLM GPT-4o mini compares the reference and target captions and writes a concise modification text, guided by instructions and two in-context examples. On the model side, FashionBLIP-2 keeps BLIP-2's ViT-G image encoder frozen, adds a bottleneck adapter $\operatorname{Adapter}(x) = x + W_u(\sigma(W_d x))$ with a downsampling factor of 16 inside each transformer layer, and replaces global-vector matching with multi-head query-candidate matching: the 32 Q-Former embeddings are token-mixed to $n_t = 12$ vectors and channel-mixed to dimension 256, and the final score sums per-head cosine similarities. This combination is what lets the model use the detailed modification texts for fine-grained retrieval.

What would settle it

Independently annotate a fresh random sample of several hundred FACap triplets without using the paper's pipeline, and check whether each generated modification text describes a real, target-specific visual change; if the faithfulness score drops below the reported 4.40 or approaches the 3.03 detail score of FashionIQ, the fine-grained gains could be an artifact of shared annotation bias between the training set and the enhFashionIQ evaluation.

Watch

Extended reading notes

Core claim

The central claim, stated as the authors would state it, is that a large-scale, automatically annotated CIR dataset in the fashion domain is both feasible and effective. FACap contains 227,680 triplets with modification texts averaging 23.38 words and a vocabulary of 9,273 words, built from Fashion200k and DeepFashion-MultiModal images paired by CLIP similarity within the same category. A quality check on 216 sampled triplets gives FACap faithfulness 4.40 (FashionIQ: 4.48), details 4.09 (FashionIQ: 3.03), and saliency 4.29 (FashionIQ: 3.60). FashionBLIP-2 freezes BLIP-2's image encoder, inserts bottleneck adapters, uses a Q-Former conditioned on the modification text, and computes similarity as a sum of cosine similarities over multiple token heads. Without fine-tuning on FashionIQ it reaches 44.63 average recall, beating the zero-shot LDRE baseline at 43.97; with fine-tuning it reaches 65.97, second to UniFashion at 67.93. Pretraining on FACap also improves SPRC and UniFashion on the re-annotated enhFashionIQ evaluation, with FashionBLIP-2 jumping from 80.32 to 87.93.

Load-bearing premise

The load-bearing premise is that the VLM and LLM produce modification texts that faithfully describe true visible differences between paired images across all 227,680 triplets, not merely the 216 triplets manually spot-checked in the quality evaluation.

Editorial extensions

If this is right

  • Automatic two-stage annotation can replace expensive specialist labeling for fashion CIR: FACap has more than twelve times as many triplets as FashionIQ with comparable faithfulness and higher detail and saliency scores.
  • Pretraining on FACap transfers to multiple architectures: SPRC, UniFashion, and FashionBLIP-2 all improve when FACap pretraining precedes FashionIQ fine-tuning, with the largest gains on the fine-grained enhFashionIQ set (UniFashion from 81.97 to 87.30; FashionBLIP-2 from 80.32 to 87.93).
  • A model trained only on FACap, with no FashionIQ fine-tuning, reaches 44.63 average recall and beats the best zero-shot baseline (LDRE, 43.97), so the dataset alone provides meaningful fashion-domain adaptation.
  • Multi-head query-candidate matching contributes independently of the adapters: in the ablations it produces a larger gain on the fine-grained evaluation than the adapter alone, and it is the component that best exploits FACap's detailed annotations.

Reading between the lines

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

  • Beyond the paper: the annotation pipeline is not fashion-specific, so the same two-stage recipe should produce usable CIR datasets for other product domains such as home decor or beauty; a direct test is to build them and measure whether retrieval gains repeat.
  • Beyond the paper: because enhFashionIQ is annotated by the same pipeline that created FACap, the size of the true fine-grained gain is best confirmed with an independent human-annotated test set; that comparison would separate shared annotation bias from genuine retrieval quality.
  • Beyond the paper: the scaling curve shows diminishing returns past half of FACap, so further gains likely require harder negative pairs or diversity-aware sampling rather than simply adding more web-crawled images.
  • Beyond the paper: the pattern across methods suggests architectures with token-level or multi-head matching profit more from detailed auto-captioned data than global-embedding models; this predicts that future fine-grained matching models will be the main beneficiaries of large auto-annotated CIR datasets.
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

4 major / 3 minor

Summary. The paper introduces FACap, a large-scale automatically constructed fashion composed image retrieval (CIR) dataset with 227,680 triplets, built from Fashion200k and DeepFashion-MultiModal images using a two-stage pipeline (InternVL image captioning followed by GPT-4o-mini modification-text generation). It also proposes FashionBLIP-2, which adapts BLIP-2 with lightweight adapters and a multi-head query-candidate matching module, and evaluates it on FashionIQ and a newly constructed enhFashionIQ benchmark under both fine-tuning and no-fine-tuning settings. The paper claims that FACap pretraining improves fashion CIR performance, especially for fine-grained modification texts, and that FashionBLIP-2 achieves state-of-the-art or competitive results.

Significance. If the claims hold, FACap would be a valuable resource: it is more than an order of magnitude larger than FashionIQ, is intended for public release, and the two-stage annotation pipeline offers a practical recipe for scaling CIR data. The experimental design is largely sound: Table 4 shows consistent gains for FashionBLIP-2 on the human-annotated FashionIQ split (+1.51 average Recall), Table 5 cleanly isolates the contributions of the adapter, multi-head matching, and the auxiliary CTR task, and Figure 5 shows scaling with dataset size. The release of code and the careful exclusion of downstream images from the pretraining pool are strengths. However, the central fine-grained claim rests almost entirely on enhFashionIQ, which is generated by the same pipeline as the training data; this circularity must be addressed before the significance can be fully credited.

major comments (4)
  1. [Section 5.1; Table 4; Table 5] The enhFashionIQ benchmark is constructed by applying the authors' own InternVL + GPT-4o-mini annotation process to FashionIQ validation images, which is the same pipeline used to build FACap. Consequently, models pretrained on FACap and evaluated on enhFashionIQ share annotation vocabulary, style, and pairing heuristics with their training data. This appears directly in the results: Table 4 shows that FACap pretraining improves FashionBLIP-2 on enhFashionIQ by +7.61 average Recall but only +1.51 on the human-annotated FashionIQ split, and UniFashion improves by +5.33 on enhFashionIQ while declining by -0.83 on FashionIQ. Table 5 shows that FashionBLIP-2 trained only on FACap (without FashionIQ fine-tuning) reaches 89.45 on enhFashionIQ, while fine-tuning on FashionIQ lowers that score to 86.93; the fact that a model that has never seen FashionIQ can outperform a FashionIQ-fine-tuned model on a FashionIQ-derived benchmark indicates that the benchmark's structure overlaps with FACap's annotations. Because the paper's central claim is that FACap is especially beneficial for fine-grained modification texts, this distribution overlap is a load-bearing threat to validity. The authors should either evaluate on an independently human-annotated fine-grained benchmark, or provide a quantitative analysis of annotation distribution overlap (e.g., retrieval tests using generated texts on human-annotated pairs) to rule out this explanation.
  2. [Section 3.2; Table 2] The quality audit samples only 216 triplets per dataset and reports mean scores with standard deviations for faithfulness, details, and saliency, but it does not report inter-annotator agreement (e.g., Fleiss' kappa or Krippendorff's alpha) or any significance test comparing FACap to FashionIQ. The claim that 'our automatically constructed dataset exhibits even higher quality' rests on differences that may be within inter-annotator noise: for faithfulness, FACap's 4.40±0.60 is actually below FashionIQ's 4.48±0.64, and the text itself acknowledges 'similar faithfulness values.' Without agreement metrics and a statistical comparison, the quality advantage is not established, and this matters because the dataset's pretraining value depends on annotation faithfulness and discriminability.
  3. [Section 4.3, Eqs. (2)-(4)] The text states 'we use the same parameters W_tm = W_cm to process x_q and x_c,' but Eq. (2) defines W_tm ∈ R^{n_t×n_q} while Eq. (3) defines W_cm ∈ R^{d_q×d_c}; these matrices have incompatible shapes and cannot be equal. Please clarify whether the sharing is across the query and candidate branches (i.e., W_tm and W_cm are each shared between x_q and x_c) or something else. As written, the matching module is not fully specified and cannot be reproduced.
  4. [Section 5.3, Table 4] The text claims 'almost all models benefit from pretraining on FACap, improving performance on the two evaluation datasets.' On the human-annotated FashionIQ split, however, SPRC improves by only 0.02 points (64.85 to 64.87) and UniFashion declines by 0.83 points (65.34 to 64.51). Only FashionBLIP-2 shows a meaningful improvement on FashionIQ (+1.51). The claim should be revised to accurately reflect that the clear gains appear on enhFashionIQ and for FashionBLIP-2, and the discussion should address why FACap pretraining does not help existing models on the human benchmark.
minor comments (3)
  1. [Section 5.1] The 'without fine-tuning' setting is not a zero-shot setting because FashionBLIP-2 is trained on FACap; the comparison against zero-shot methods should be explicitly framed as supervised pretraining versus zero-shot to avoid confusion.
  2. [Figure 5] Figure 5 is described as results on Fashion IQ without fine-tuning, but the text says performance on 'both evaluation benchmarks'; please clarify in the caption which curve corresponds to which benchmark.
  3. [Section 4.4, Eq. (5)] The notation for the contrastive loss omits the modification text in the query representation; consider making explicit that the query embedding x_i is a function of both the reference image and the modification text.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity: the fine-grained gains are measured on enhFashionIQ, which is produced by the same VLM/LLM pipeline as the FACap training set.

  1. fitted input called prediction [Section 5.1 (enhFashionIQ construction), Table 4, Table 5; abstract claim]
    "However, since the annotations in Fashion IQ are noisy as shown in Figure 1 and Table 2, we enhance its quality by applying our automatic annotation process to the images from the Fashion IQ validation split. We create a triplet for each unique image in the validation split and generate a total of 15,536 CIR triplets, which we name enhFashionIQ, for fine-grained CIR evaluation."

    FACap and enhFashionIQ are both outputs of the same two-stage annotation pipeline described in Section 3.1: InternVL captioning, GPT-4o-mini modification-text generation, and CLIP-similarity image pairing. The headline fine-grained claim is then measured on enhFashionIQ: Table 4 credits FACap pretraining with raising FashionBLIP-2's enhFashionIQ average Recall from 80.32 to 87.93, while the human-annotated FashionIQ gain is only 64.46 to 65.97; Table 5 even shows the FACap-only model scoring 89.45 on enhFashionIQ, above the FashionIQ-fine-tuned model.

full rationale

The model equations and training losses in Section 4 are not circular: FashionBLIP-2 is trained with standard contrastive CIR and CTR losses and compared against external zero-shot and fine-tuned baselines on the human-annotated FashionIQ validation set. There is no load-bearing self-citation chain or imported uniqueness theorem; citations to BLIP-2, InternVL, and VL-Adapter are external and do not substitute for evidence. The main circularity is benchmark-level: enhFashionIQ is generated by the same VLM/LLM pipeline used to build FACap, and the strongest 'especially fine-grained' result rests on that benchmark, so part of the measured gain is a fit to the annotation pipeline's own distribution. This is a partial circularity, not a definitional equivalence; the improvement on human-annotated FashionIQ (64.46 to 65.97 in Table 4) provides independent evidence for the central transfer claim.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a data-construction pipeline and a set of hyperparameters, not on new physical or theoretical entities. The most important unverified load is the quality and consistency of the automatic captions and modification texts; the paper's quality audit is small (216 triplets) and the enhFashionIQ evaluation set is built by the same pipeline.

free parameters (5)
  • caption_max_tokens = 128
    Chosen in Section 3.1 to balance caption detail against hallucination risk in InternVL generation; directly controls the information available to the LLM difference generator.
  • CLIP_similarity_top_k = 20
    Pairs are randomly drawn from the top-20 CLIP-similar same-category images (Section 3.1); this threshold sets the difficulty range of the modification text and is not derived from a principled criterion.
  • adapter_bottleneck_ratio = 16
    Downsampling factor for the adapters in Section 5.1; determines how much fashion-specific capacity the frozen BLIP-2 image encoder can learn.
  • q_former_query_tokens = 32
    Number of learnable queries in the Q-Former (Section 5.1); sets the token budget from which the multi-head matcher draws its representations.
  • matching_heads_and_dim = n_t=12, d_c=256
    Number of retained tokens after token mixing and the channel dimension after channel mixing (Section 5.1); both are hand-chosen and control the granularity of matching.
assumptions (4)
  • domain assumption CLIP cosine similarity ranks pairs in a useful range for CIR modification texts.
    Section 3.1 'Image pairing' uses CLIP features to pick top-20 neighbors; no evidence that this similarity range produces modifications that are neither too global nor too subtle.
  • domain assumption InternVL captions are detailed and hallucination-limited enough for difference generation.
    Section 3.1 first stage assumes 128-token InternVL captions plus noisy product context yield accurate descriptions; only spot-checked indirectly in Section 3.2.
  • domain assumption GPT-4o mini's modification texts are faithful, detailed, and salient across the full dataset.
    Section 3.1 second stage relies on an LLM to compress two captions into a retrieval-relevant sentence; the quality check covers 216 of 227,680 triplets with no inter-annotator agreement.
  • domain assumption Frozen BLIP-2 image encoder plus residual adapters preserves general knowledge while learning fashion details.
    Section 4.2 assumes the adapter residual pathway improves fine-grained fashion features without catastrophic forgetting; supported only by ablations, not by analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FACap: A Large-scale Fashion Dataset for Fine-grained Composed Image Retrieval." pith.science (2026). https://pith.science/paper/F7EB7Y7Z

@misc{pith2026250707135,
  author       = {Pith},
  title        = {Pith review of: FACap: A Large-scale Fashion Dataset for Fine-grained Composed Image Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F7EB7Y7Z}},
  note         = {Machine review of arXiv:2507.07135}
}
read the original abstract

The composed image retrieval (CIR) task is to retrieve target images given a reference image and a modification text. Recent methods for CIR leverage large pretrained vision-language models (VLMs) and achieve good performance on general-domain concepts like color and texture. However, they still struggle with application domains like fashion, because the rich and diverse vocabulary used in fashion requires specific fine-grained vision and language understanding. An additional difficulty is the lack of large-scale fashion datasets with detailed and relevant annotations, due to the expensive cost of manual annotation by specialists. To address these challenges, we introduce FACap, a large-scale, automatically constructed fashion-domain CIR dataset. It leverages web-sourced fashion images and a two-stage annotation pipeline powered by a VLM and a large language model (LLM) to generate accurate and detailed modification texts. Then, we propose a new CIR model FashionBLIP-2, which fine-tunes the general-domain BLIP-2 model on FACap with lightweight adapters and multi-head query-candidate matching to better account for fine-grained fashion-specific information. FashionBLIP-2 is evaluated with and without additional fine-tuning on the Fashion IQ benchmark and the enhanced evaluation dataset enhFashionIQ, leveraging our pipeline to obtain higher-quality annotations. Experimental results show that the combination of FashionBLIP-2 and pretraining with FACap significantly improves the model's performance in fashion CIR especially for retrieval with fine-grained modification texts, demonstrating the value of our dataset and approach in a highly demanding environment such as e-commerce websites. Code is available at https://fgxaos.github.io/facap-paper-website/.

Figures

Figures reproduced from arXiv: 2507.07135 by the authors.

Figure 1
Figure 1. Our automatically constructed FACap dataset offers more detailed and accurate annotations than existing datasets for the fashion CIR task. the crowdsourced captions in Fashion IQ are short, noisy and lack details, as shown in Figure 1a and confirmed by our quality evaluation in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The proposed data construction pipeline to automatically generate CIR triplets. The images are extracted from large existing fashion datasets, then paired based on their visual similarity with images from the same product category. Then, our two-stage annotation process captions the images with a VLM, and an LLM generates a synthetic description of the changes applied on the reference image to obtain the target imag… view at source ↗
Figure 3
Figure 3. Examples from the FACap dataset. The caption of each image pair corresponds to their modification text. issues with context length. To address this, various image token compression techniques [48, 49] have been proposed for VLMs. Second, on the data side, multi-view image datasets [45] remain limited, restricting the ability of current VLMs to excel in multi-image reasoning tasks. In this work, instead of directly u… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Overview of the FashionBLIP-2 model. Left: The input images are encoded using a pretrained image encoder with adapter modules, and further processed by a Q-Former module. The similarity between the two obtained representations is computed using multi-head query-candida…
Figure 5
Figure 5. Figure 5: Results on Fashion IQ dataset without fine-tuning using different sizes of the FACap dataset [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results of FashionBLIP-2 on Fashion IQ (rows 1 and 3) and enhFashionIQ (row 2). The rank of the ground-truth image (framed in green) among the retrieved results is specified on the right. progressively larger subsets of FACap and evaluate the resulting mode…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 39 canonical work pages

  1. [1]

    H. Wu, Y. Gao, X. Guo, Z. Al-Halah, S. Rennie, K. Grauman, R. Feris, 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, 2021, pp. 11307–11317

    work page 2021

  2. [2]

    S. Shao, K. Chen, A. Karpur, Q. Cui, A. Araujo, B. Cao, Global features are all you need for image retrieval and reranking, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 11036–11046

    work page 2023

  3. [3]

    J. Rao, F. Wang, L. Ding, S. Qi, Y. Zhan, W. Liu, D. Tao, Where does the performance improvement come from? -a reproducibility concern about image-text retrieval, in: Proceedings of the 45th international ACM SIGIR conference on research and development in information retrieval, 2022, pp. 2727–2737

    work page 2022

  4. [4]

    N. Vo, L. Jiang, C. Sun, K. Murphy, L.-J. Li, L. Fei-Fei, J. Hays, Composing text and image for image retrieval-an empirical odyssey, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 6439–6448

    work page 2019

  5. [5]

    Baldrati, M

    A. Baldrati, M. Bertini, T. Uricchio, A. Del Bimbo, Conditioned and composed image retrieval com- bining and partially fine-tuning clip-based features, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 4959–4968

    work page 2022

  6. [6]

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

    work page 2024

  7. [7]

    Z. Feng, R. Zhang, Z. Nie, Improving composed image retrieval via contrastive learning with scaling positives and negatives, in: Proceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 1632–1641

    work page 2024

  8. [8]

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

    work page 2024

Show all 56 references
  1. [9]

    Y. Zhao, Y. Song, Q. Jin, 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, 2022, pp. 1012–1021

    2022

  2. [10]

    X. Zhao, Y. Zhang, W. Zhang, X.-M. Wu, UniFashion: A unified vision-language model for multimodal fashion retrieval and generation, in: Y. Al-Onaizan, M. Bansal, Y.-N. Chen (Eds.), Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Associat...

    2024 doi

  3. [11]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al., Learning transferable visual models from natural language supervision, in: International conference on machine learning, PMLR, 2021, pp. 8748–8763

    2021

  4. [12]

    J. Li, D. Li, S. Savarese, S. Hoi, Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models, in: International conference on machine learning, PMLR, 2023, pp. 19730–19742

    2023

  5. [13]

    D. H. Park, T. Darrell, A. Rohrbach, Robust change captioning, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 4624–4633

    2019

  6. [14]

    Baldrati, L

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

    2023

  7. [15]

    Saito, K

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

    2023

  8. [16]

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

    2024

  9. [17]

    Z. Yang, D. Xue, S. Qian, W. Dong, C. Xu, 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, 2024, pp. 80–90

    2024

  10. [18]

    H. Lin, H. Wen, X. Song, M. Liu, Y. Hu, L. Nie, Fine-grained textual inversion network for zero-shot composed image retrieval, in: Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2024, pp. 240–250

    2024

  11. [19]

    Karthik, K

    S. Karthik, K. Roth, M. Mancini, Z. Akata, Vision-by-language for training-free compositional image retrieval, arXiv preprint arXiv:2310.09291 (2023)

    2023 arXiv

  12. [20]

    Y. Bai, X. Xu, Y. Liu, S. Khan, F. Khan, W. Zuo, R. S. M. Goh, C.-M. Feng, et al., Sentence-level prompts benefit composed image retrieval, in: The Twelfth International Conference on Learning Representations, 2024

    2024

  13. [21]

    J. Li, D. Li, C. Xiong, S. Hoi, Blip: Bootstrapping language-image pre-training for unified vision- language understanding and generation, in: International conference on machine learning, PMLR, 2022, pp. 12888–12900

    2022

  14. [22]

    H. Wen, X. Zhang, X. Song, Y. Wei, L. Nie, Target-guided composed image retrieval, in: Proceedings of the 31st ACM International Conference on Multimedia, 2023, pp. 915–923

    2023

  15. [23]

    Delmas, R

    G. Delmas, R. S. de Rezende, G. Csurka, D. Larlus, Artemis: Attention-based retrieval with text-explicit matching and implicit similarity, ICLR (2022)

    2022

  16. [24]

    Jiang, Y

    X. Jiang, Y. Wang, M. Li, Y. Wu, B. Hu, X. Qian, Cala: Complementary association learning for augmenting comoposed image retrieval, in: Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2024, pp. 2177–2187

    2024

  17. [25]

    Z. Liu, C. Rodriguez-Opazo, D. Teney, S. Gould, Image retrieval on real-life images with pre- trained vision-and-language models, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 2125–2134

    2021

  18. [26]

    T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, C. L. Zitnick, Microsoft coco: Common objects in context, in: Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, Springer, 2014, pp...

    2014

  19. [27]

    Schuhmann, R

    C. Schuhmann, R. Beaumont, R. Vencu, C. Gordon, R. Wightman, M. Cherti, T. Coombes, A. Katta, C. Mullis, M. Wortsman, et al., Laion-5b: An open large-scale dataset for training next generation image-text models, Advances in Neural Information Processing Systems 35 (2022) 25278–25294

    2022

  20. [28]

    R. Gal, Y. Alaluf, Y. Atzmon, O. Patashnik, A. H. Bermano, G. Chechik, D. Cohen-or, An image is worth one word: Personalizing text-to-image generation using textual inversion, in: The Eleventh International Conference on Learning Representations, ????

  21. [29]

    G. Gu, S. Chun, W. Kim, H. Jun, Y. Kang, S. Yun, Compodiff: Versatile composed image retrieval with latent diffusion, arXiv preprint arXiv:2303.11916 (2023)

    2023 arXiv

  22. [30]

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

    2024

  23. [31]

    Ventura, A

    L. Ventura, A. Yang, C. Schmid, G. Varol, Covr-2: Automatic data construction for composed video retrieval, IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    2024

  24. [32]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, B. Ommer, High-resolution image synthesis with latent diffusion models, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10684–10695

    2022

  25. [33]

    Hertz, R

    A. Hertz, R. Mokady, J. Tenenbaum, K. Aberman, Y. Pritch, D. Cohen-Or, Prompt-to- prompt image editing with cross attention control, 2022. URL: https://arxiv.org/abs/2208.01626. arXiv:2208.01626

    2022 arXiv

  26. [34]

    X. Han, L. Yu, X. Zhu, L. Zhang, Y.-Z. Song, T. Xiang, Fashionvil: Fashion-focused vision-and- language representation learning, in: European conference on computer vision, Springer, 2022, pp. 634–651

    2022

  27. [35]

    X. Han, X. Zhu, L. Yu, L. Zhang, Y.-Z. Song, T. Xiang, Fame-vil: Multi-tasking vision-language model for heterogeneous fashion tasks, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 2669–2680

    2023

  28. [36]

    Brown, B

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al., Language models are few-shot learners, Advances in neural information processing systems 33 (2020) 1877–1901

    2020

  29. [37]

    Dubey, A

    A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Yang, A. Fan, et al., The llama 3 herd of models, arXiv preprint arXiv:2407.21783 (2024)

    2024 arXiv

  30. [38]

    H. Liu, C. Li, Q. Wu, Y. J. Lee, Visual instruction tuning, Advances in neural information processing systems 36 (2024)

    2024

  31. [39]

    Achiam, S

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat, et al., Gpt-4 technical report, arXiv preprint arXiv:2303.08774 (2023)

    2023 arXiv

  32. [40]

    Z. Chen, J. Wu, W. Wang, W. Su, G. Chen, S. Xing, M. Zhong, Q. Zhang, X. Zhu, L. Lu, et al., Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2...

    2024

  33. [41]

    J. Wang, Z. Yang, X. Hu, L. Li, K. Lin, Z. Gan, Z. Liu, C. Liu, L. Wang, Git: A generative image-to-text transformer for vision and language, arXiv preprint arXiv:2205.14100 (2022)

    2022 arXiv

  34. [42]

    Z. Gan, L. Li, C. Li, L. Wang, Z. Liu, J. Gao, et al., Vision-language pre-training: Basics, recent advances, and future trends, Foundations and Trends® in Computer Graphics and Vision 14 (2022) 163–352

    2022

  35. [43]

    Antol, A

    S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra, C. L. Zitnick, D. Parikh, Vqa: Visual question answering, in: Proceedings of the IEEE international conference on computer vision, 2015, pp. 2425–2433

    2015

  36. [44]

    X. Yang, H. Zhang, D. Jin, Y. Liu, C.-H. Wu, J. Tan, D. Xie, J. Wang, X. Wang, Fashion captioning: Towards generating accurate descriptions with semantic rewards, in: ECCV, 2020

    2020

  37. [45]

    H. Liu, C. Li, Y. Li, B. Li, Y. Zhang, S. Shen, Y. J. Lee, Llava-next: Improved reasoning, ocr, and world knowledge, 2024. URL: https://llava-vl.github.io/blog/2024-01-30-llava-next/

    2024

  38. [46]

    B. Li, Y. Zhang, D. Guo, R. Zhang, F. Li, H. Zhang, K. Zhang, P. Zhang, Y. Li, Z. Liu, et al., Llava-onevision: Easy visual task transfer, arXiv preprint arXiv:2408.03326 (2024)

    2024 arXiv

  39. [47]

    F. Li, R. Zhang, H. Zhang, Y. Zhang, B. Li, W. Li, Z. Ma, C. Li, Llava-next-interleave: Tackling multi-image, video, and 3d in large multimodal models, arXiv preprint arXiv:2407.07895 (2024)

    2024 arXiv

  40. [48]

    X. Shen, Y. Xiong, C. Zhao, L. Wu, J. Chen, C. Zhu, Z. Liu, F. Xiao, B. Varadarajan, F. Bordes, et al., Longvu: Spatiotemporal adaptive compression for long video-language understanding, arXiv preprint arXiv:2410.17434 (2024)

    2024 arXiv

  41. [49]

    W. Li, Y. Yuan, J. Liu, D. Tang, S. Wang, J. Qin, J. Zhu, L. Zhang, Tokenpacker: Efficient visual projector for multimodal llm, arXiv preprint arXiv:2407.02392 (2024)

    2024 arXiv

  42. [50]

    X. Han, Z. Wu, P. X. Huang, X. Zhang, M. Zhu, Y. Li, Y. Zhao, L. S. Davis, Automatic spatially-aware fashion concept discovery, in: Proceedings of the IEEE international conference on computer vision, 2017, pp. 1463–1471

    2017

  43. [51]

    Jiang, S

    Y. Jiang, S. Yang, H. Qiu, W. Wu, C. C. Loy, Z. Liu, Text2human: Text-driven controllable human image generation, ACM Transactions on Graphics (TOG) 41 (2022) 1–11. doi: 10.1145/3528223. 3530104

    2022 doi

  44. [52]

    Z. Liu, P. Luo, S. Qiu, X. Wang, X. Tang, Deepfashion: Powering robust clothes recognition and retrieval with rich annotations, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 1096–1104

    2016

  45. [53]

    Y.-L. Sung, J. Cho, M. Bansal, Vl-adapter: Parameter-efficient transfer learning for vision-and- language tasks, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 5227–5237

    2022

  46. [54]

    Waswani, N

    A. Waswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. Gomez, L. Kaiser, I. Polosukhin, Attention is all you need, in: NIPS, 2017

    2017

  47. [55]

    Hendrycks, K

    D. Hendrycks, K. Gimpel, Gaussian error linear units (gelus), arXiv preprint arXiv:1606.08415 (2016)

    2016 arXiv

  48. [56]

    I. O. Tolstikhin, N. Houlsby, A. Kolesnikov, L. Beyer, X. Zhai, T. Unterthiner, J. Yung, A. Steiner, D. Keysers, J. Uszkoreit, et al., Mlp-mixer: An all-mlp architecture for vision, Advances in neural information processing systems 34 (2021) 24261–24272

    2021

Pith tools

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