Pith. sign in

REVIEW 3 major objections 6 minor 50 references

AttriPrompt: Dynamic Prompt Composition Learning for CLIP

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper claims that dynamic, image-conditioned text prompts improve CLIP's few-shot and zero-shot generalization.

desk verdict Solid empirical prompt-learning paper with a genuinely new retrieval mechanism, but the discrete routing gradient path is under-specified and reproducibility is incomplete. read the letter →

arxiv 2509.05949 v1 pith:QBUECWOV submitted 2025-09-07 cs.CV

Qiqi Zhan , Shiwei Li , Qingjie Liu , Yunhong Wang This is my paper
classification cs.CV
keywords transferlearningvision-languagemodelspromptCLIPadaptationbase-to-novelgeneralizationcross-datasetdomaindynamiccomposition
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 claims that CLIP's text prompts can be made both deeper and content-aware by pulling visual attributes out of the vision encoder's intermediate layers. In AttriPrompt, each input image retrieves its own set of prompts from a learned pool, the retrieved prompts are inserted into every layer of the text encoder, and three auxiliary losses preserve the original model's knowledge while sharpening image-text alignment. Across 11 recognition datasets, the method reports the best average base-to-novel harmonic mean (81.09), including a 7.37-point gain on novel EuroSAT classes over the strongest compared method. If right, this is a practical route to transfer CLIP to new domains with less data and less forgetting.

What carries the argument

The load-bearing object is the Attribute Retrieval module: a prompt pool of M learnable prompts, each paired with a key vector; a per-layer clustering step (k-means) that reduces a layer's visual tokens to k attribute features; and a cosine-similarity top-1 selection that picks a prompt for each attribute, which is then concatenated into the corresponding text-encoder layer. Around it, Dual-stream Contrastive Learning adds channel-wise scaling/shifting heads on both encoders and re-applies the contrastive loss in that transformed space, while Self-Regularization imposes an L1 penalty between prompted and non-prompted text features. Together they let the text side carry visual detail, keep th

What would settle it

A reader could settle the dynamic-routing claim by replacing the top-1 retrieval in Eq. 3 with random selection from the same prompt pool, keeping all other losses and hyperparameters fixed. If the harmonic mean stays within roughly one point of 81.09, the content-aware selection is not doing the claimed work. A second check is to verify that gradients actually reach the keys: the paper gives no estimator for the discrete top-1 operation, so the routing may be shaped only by the small L_match term.

Watch

Extended reading notes

Core claim

AttriPrompt's central proposal is to make text prompts dynamic rather than static. Instead of sharing one learned prompt string across all inputs, the model clusters the visual tokens of each intermediate vision layer into k attribute features, uses those features to query a pool of M learnable prompts by cosine similarity, and concatenates the top-1 prompt for each attribute into the corresponding text-encoder layer. The prompted text features are then trained with three complementary objectives: the standard CLIP contrastive loss, a dual-stream contrastive loss computed in a channel-wise transformed space, and an L1 self-regularization that keeps the prompted features close to the frozen C

Load-bearing premise

The approach assumes the top-1 prompt choice in Eq. 3 can be learned through gradient descent even though the paper specifies no gradient path or estimator for that discrete selection; if the routing does not actually learn, the method reduces to a fixed prompt pool plus regularization.

Editorial extensions

If this is right

  • On 11 base-to-novel datasets, the reported average harmonic mean is 81.09, versus 79.97 for PromptSRC and 78.55 for MaPLe.
  • The method reports a 7.37-point improvement on EuroSAT's novel classes, from 73.90 to 81.27.
  • After ImageNet training, it reaches 67.17 average accuracy on ten unseen datasets, beating PromptSRC on 9 of 10.
  • On ImageNet distribution-shift sets, average accuracy is 61.23, higher than PromptSRC's 60.65 and every compared baseline.
  • In a 4-shot base-to-novel setting, it matches PromptSRC's 16-shot accuracy, implying a 4-fold data-efficiency gain.

Reading between the lines

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

  • Editorial inference: Since the retrieval module only touches intermediate features and a prompt pool, it could be grafted onto other frozen vision-language encoders or other prompt-learning methods; the paper does not test this, so this is an extension.
  • Editorial inference: The large EuroSAT novel-class gain suggests the method should be tried on other domains where class names are weak and visual attributes matter, such as satellite, aerial, or medical imaging.
  • Editorial inference: The paper's dynamic-routing claim would be more convincing with an analysis of which prompts are selected for which classes; that verification is absent, but a reader could produce it by logging the argmax selections on the test set.
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 / 6 minor

Summary. The paper proposes AttriPrompt, a prompt-learning approach for CLIP that aims to make text prompts content-adaptive and to inject fine-grained visual information into the text encoder. The method has three main components: (1) an Attribute Retrieval module that clusters intermediate-layer visual features and, for each cluster, retrieves a top-1 prompt from a shared prompt pool using cosine similarity (Eq. 3); (2) a Dual-stream Contrastive Learning loss that aligns transformed image/text features through channel-wise heads; and (3) a Self-Regularization L1 loss between prompted and non-prompted text features, along with prompt-diversity and retrieval-matching losses. The method is evaluated on base-to-novel generalization, cross-dataset transfer, and domain generalization using ViT-B/16. The paper reports an average harmonic mean of 81.09 across 11 datasets in the base-to-novel setting, compared with PromptSRC's 79.97, and highlights a 7.37% improvement on EuroSAT novel classes.

Significance. If the underlying retrieval mechanism actually learns to route prompts based on image content, the paper proposes a genuinely novel direction for prompt learning: using hierarchical visual attributes to dynamically compose text prompts. The idea is timely and the broad empirical evaluation across three benchmarks is a strength. The component ablations are consistent with each added loss improving performance, and the compute-cost analysis is useful. However, the significance is conditional on resolving a load-bearing training issue: the paper never specifies how gradients flow through the discrete top-1 retrieval or through the k-means clustering. Without this, the 'dynamic prompt' claim is not substantiated, and the reported improvements could be attributed to the extra prompt parameters and auxiliary losses rather than to learned content-aware routing. The absence of error bars and code also makes it hard to interpret the small average margins over strong baselines.

major comments (3)
  1. [§4.2, Eq. (3); §4.5, Eq. (9)] The top-1 prompt retrieval is non-differentiable, yet no gradient estimator is described. After computing softmax similarities in Eq. (3), the paper selects the highest-scoring key and concatenates the corresponding prompt; it does not specify a straight-through estimator, Gumbel-softmax, REINFORCE-style gradient, or detached routing. The only explicit coupling between keys and image features is L_match in Eq. (9), where the key index j is never defined and the weight is only λ4=0.01. Furthermore, the k-means clustering used to obtain attribute features is non-differentiable, and the paper does not state whether centroids are treated as constants. As written, the routing may be effectively random or fixed, making the method a fixed prompt pool plus regularization. The +2.58 HM gain attributed to Attribute Retrieval in Table 4 could then come from added prompt capacity and the L_div/L_mat
  2. [§5.1, Figures 3/4/6/7, Table 6] Key hyperparameters (M, k, λ1–λ4, Lp) are selected through ablations on the same benchmark datasets used in the headline comparisons, with no held-out validation split and no reported error bars. Given that the average HM gain over PromptSRC is only 1.12 points (81.09 vs. 79.97) and that on some datasets the method is not the best (e.g., Food101 HM 91.35 vs. MaPLe 91.38; FGVCAircraft HM 39.80 vs. PromptSRC 40.15), the 'state-of-the-art' claim requires significance testing or at least per-seed standard deviations. Please clarify whether the hyperparameters were fixed before seeing test results and report variance over the three runs.
  3. [§5.2, Table 1] The text states that AttriPrompt 'surpasses existing approaches in performance on base categories, novel categories, and the harmonic mean.' This is not supported for every dataset in Table 1: Food101 and FGVCAircraft are counterexamples. The sentence should be qualified as applying to the average across 11 datasets, or per-dataset comparisons should be tested for significance.
minor comments (6)
  1. [Eq. (9)] The index j in L_match is undefined; it presumably refers to the retrieved key for attribute a_l_i, but this should be stated explicitly.
  2. [Table 3] The target average is reported as 61.24% in the text but 61.23% in the table; the ImageNet-Sketch value is written as 49.6 rather than 49.60.
  3. [Abstract] 'constrastive' should be 'contrastive'. Also, 'CoOP'/'CoCoOP' are inconsistently capitalized; use the standard 'CoOp'/'CoCoOp'.
  4. [Appendix B] The complexity analysis says k-means is applied to 'n×n image features' and gives O(n^2 d). For ViT patch tokens, the input is N tokens of dimension d, not an n×n feature map; the complexity should be stated in terms of N and the k-means iterations.
  5. [Figures 3–7] Axis labels and legends are not visible in the provided text; please ensure the final figures clearly label the ablated hyperparameters and the metric being plotted.
  6. [General] No code is provided, and the seed/run details for the three reported averages are not given. Releasing code and seeds would substantially improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical, evaluated on held-out classes; no equation-level reduction or load-bearing self-citation is present.

full rationale

AttriPrompt's central claims are benchmark results on base-to-novel, cross-dataset, and domain generalization. These are measured on novel/unseen classes that are not used in fitting the prompt pool, keys, or heads, so the numbers are not forced by construction. The Attribute Retrieval module (Eq. 3) selects prompts by cosine similarity between clustered visual features and learnable keys; although the top-1 selection is discrete and the paper does not specify a gradient estimator, that is a trainability/correctness issue, not a circularity, because the routing is not defined in terms of the reported accuracies. The Self-Regularization loss L_cc (Eq. 5) is explicitly adopted from PromptSRC ([20]); this is a borrowed component cited to prior work, not a self-citation, and it is an auxiliary regularizer rather than the source of the claimed novelty. The Dual-stream Contrastive Learning and L_div losses are independent extra objectives. No self-citation chain, uniqueness theorem, or renamed known result is load-bearing. The only noted weaknesses—missing gradient-path specification for discrete retrieval and k-means clustering—concern whether the method trains as intended, which is a correctness risk, not a circular reduction.

Assumptions & free parameters 7 free parameters · 5 assumptions · 2 invented entities

The central method relies on many tuned hyperparameters and domain assumptions about CLIP features. These are standard for empirical prompt learning, but several, especially gradient flow through discrete retrieval, are not verified.

free parameters (7)
  • lambda1 (dual-stream contrastive weight) = 0.5
    Selected by ablation in Figure 6; controls the balance between original and auxiliary contrastive predictions at training and inference.
  • lambda2 (self-regularization weight) = 25
    Set in Section 5.1; no dedicated ablation is shown, but it heavily weights the L_cc term and is a tuned constant.
  • lambda3 (prompt diversity weight) = 0.1
    Selected by sensitivity analysis on Caltech101 and DTD in Table 6.
  • lambda4 (retrieval matching weight) = 0.01
    Selected by sensitivity analysis on Caltech101 and DTD in Table 6.
  • Prompt pool size M = 12
    Chosen by ablation in Figure 3.
  • Number of attribute clusters k = 4
    Chosen by ablation in Figure 4.
  • Prompt length Lp = 4
    Chosen by ablation in Figure 7.
assumptions (5)
  • domain assumption Intermediate vision features of frozen CLIP can be clustered into semantically meaningful attributes that correspond to useful text prompts.
    Section 4.2 relies on this to justify the Attribute Retrieval module; it is not established by external evidence.
  • domain assumption Cosine similarity in CLIP embedding space is a valid measure for retrieving prompts from visual attribute features.
    Used in Eq. 3; the paper provides no independent validation that key-image cosine similarity ranks good prompts.
  • domain assumption L1 consistency between prompted and non-prompted text features preserves CLIP generalization.
    Section 4.3 follows PromptSRC [20]; it is treated as given rather than derived.
  • domain assumption k-means with 50 iterations gives stable attribute clusters in the small-data regime.
    Appendix A states this implementation choice; no stability analysis is provided.
  • ad hoc to paper The differentiable losses are sufficient to train a discrete top-1 retrieval mechanism.
    Eqs. 3-4 and Eq. 9 imply end-to-end learning through retrieval, but no gradient estimator is described; this is load-bearing for the dynamic-prompt claim.
invented entities (2)
  • Shared prompt pool with key vectors
    purpose: Stores M learnable prompt embeddings, each with a key, so image attributes can retrieve a content-specific prompt for each text layer.
    Supported only by internal ablations in Table 4 and Figure 3; no external prediction or formal guarantee is given.
  • Channel-wise transformation heads
    purpose: Map image and text features into a new aligned space for the auxiliary dual-stream contrastive loss.
    Internal ablation shows a gain, but there is no independent evidence that the learned heads capture fine-grained features as claimed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AttriPrompt: Dynamic Prompt Composition Learning for CLIP." pith.science (2026). https://pith.science/paper/QBUECWOV

@misc{pith2026250905949,
  author       = {Pith},
  title        = {Pith review of: AttriPrompt: Dynamic Prompt Composition Learning for CLIP},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QBUECWOV}},
  note         = {Machine review of arXiv:2509.05949}
}
read the original abstract

The evolution of prompt learning methodologies has driven exploration of deeper prompt designs to enhance model performance. However, current deep text prompting approaches suffer from two critical limitations: Over-reliance on constrastive learning objectives that prioritize high-level semantic alignment, neglecting fine-grained feature optimization; Static prompts across all input categories, preventing content-aware adaptation. To address these limitations, we propose AttriPrompt-a novel framework that enhances and refines textual semantic representations by leveraging the intermediate-layer features of CLIP's vision encoder. We designed an Attribute Retrieval module that first clusters visual features from each layer. The aggregated visual features retrieve semantically similar prompts from a prompt pool, which are then concatenated to the input of every layer in the text encoder. Leveraging hierarchical visual information embedded in prompted text features, we introduce Dual-stream Contrastive Learning to realize fine-grained alignment. Furthermore, we introduce a Self-Regularization mechanism by applying explicit regularization constraints between the prompted and non-prompted text features to prevent overfitting on limited training data. Extensive experiments across three benchmarks demonstrate AttriPrompt's superiority over state-of-the-art methods, achieving up to 7.37\% improvement in the base-to-novel setting. The observed strength of our method in cross-domain knowledge transfer positions vision-language pre-trained models as more viable solutions for real-world implementation.

Figures

Figures reproduced from arXiv: 2509.05949 by the authors.

Figure 1
Figure 1. Comparison of our method with other prompt learn [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The framework of AttriPrompt, which is composed of Attributes Retrieval, Self Regularization and Dual-stream [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Ablation study on prompt pool size. To verify the effectiveness of our proposed method, we con￾ducted ablation studies on the key components, with the results summarized in [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Ablation study on 𝑘. 2-shot 4-shot 8-shot 16-shot Number of Shots 74.5 75.0 75.5 76.0 76.5 77.0 77.5 78.0 Accuracy (%) 75.4% 76.5% 77.0% PromptSRC @ 16-shot 77.6% MaPLe @ 16-shot [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Comparison of few-shot image recognition among [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Ablation study on 𝜆1. Evaluation metrics. In the base-to-novel generalization setting, we evaluate performance by reporting the top-1 accuracy on both base and novel classes across all datasets. To quantify the overall generalization ability, we also compute the harmon…
Figure 7
Figure 7. Figure 7: Ablation study on 𝐿𝑝 . C More Ablation study Dual-stream Contrastive Learning hyper-parameters. Since the hyperparameter 𝜆1 can influence the model’s predictions dur￾ing inference, we conduct an ablation study to evaluate its impact. As shown in [PITH_FULL_IMAGE:figur…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 26 canonical work pages

  1. [1]

    Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. 2022. Visual prompting: Modifying pixel space to adapt pre-trained models.arXiv preprint arXiv:2203.172743, 11-12 (2022), 3

    arXiv 2022

  2. [2]

    Hanoona Bangalath, Muhammad Maaz, Muhammad Uzair Khattak, Salman H Khan, and Fahad Shahbaz Khan. 2022. Bridging the gap between object and image-level representations for open-vocabulary detection.Advances in Neural Information Processing Systems35 (2022), 33781–33794

    work page 2022

  3. [3]

    Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. 2014. Food-101–mining discriminative components with random forests. InComputer vision–ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13. Springer, 446–461

    work page 2014

  4. [4]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. InInterna- tional conference on machine learning. PmLR, 1597–1607

    2020

  5. [5]

    Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and An- drea Vedaldi. 2014. Describing textures in the wild. InProceedings of the IEEE conference on computer vision and pattern recognition. 3606–3613

    work page 2014

  6. [6]

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition. Ieee, 248–255

    2009

  7. [7]

    Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor G Turrisi da Costa, Cees GM Snoek, Georgios Tzimiropoulos, and Brais Martinez. 2023. Bayesian prompt learning for image-language model generalization. InProceed- ings of the IEEE/CVF International Conference on Computer Vision. 15237–15246

    work page 2023

  8. [8]

    Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. 2022. Decoupling zero-shot semantic segmentation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 11583–11592

    work page 2022

Show all 50 references
  1. [9]

    Ning Ding, Shengding Hu, Weilin Zhao, Yulin Chen, Zhiyuan Liu, Hai-Tao Zheng, and Maosong Sun. 2021. Openprompt: An open-source framework for prompt- learning.arXiv preprint arXiv:2111.01998(2021)

    2021 arXiv

  2. [10]

    Li Fei-Fei, Rob Fergus, and Pietro Perona. 2004. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In2004 conference on computer vision and pattern recognition workshop. IEEE, 178–178

    2004

  3. [11]

    Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. 2022. Promptdet: Towards open-vocabulary detection using uncurated images. InEuropean conference on computer vision. Springer, 701–717

    2022

  4. [12]

    Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. 2024. Clip-adapter: Better vision-language models with feature adapters.International Journal of Computer Vision132, 2 (2024), 581–595

    2024

  5. [13]

    Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. 2021. Open-vocabulary object detection via vision and language knowledge distillation.arXiv preprint arXiv:2104.13921(2021)

    2021 arXiv

  6. [14]

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

    2016

  7. [15]

    Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. 2019. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing12, 7 (2019), 2217–2226

    2019

  8. [16]

    Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. 2021. The many faces of robustness: A critical analysis of out-of-distribution generalization. InProceedings of the IEEE/CVF internatio...

    2021

  9. [17]

    Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song

  10. [18]

    Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. 2021. Scaling up visual and vision- language representation learning with noisy text supervision. InInternational conference on machine learning. PMLR, 4904–4916

    2021

  11. [19]

    Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. 2023. Maple: Multi-modal prompt learning. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition. 19113–19122

    2023

  12. [20]

    Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. 2023. Self-regulating prompts: Foundational model adaptation without forgetting. InProceedings of the IEEE/CVF international conference on computer vision. 15190–15200

    2023

  13. [21]

    Konwoo Kim, Michael Laskin, Igor Mordatch, and Deepak Pathak. 2021. How to adapt your large-scale vision-and-language model. (2021)

    2021

  14. [22]

    Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 2013. 3d object repre- sentations for fine-grained categorization. InProceedings of the IEEE international conference on computer vision workshops. 554–561

    2013

  15. [23]

    Weinberger, Serge J

    Boyi Li, Kilian Q. Weinberger, Serge J. Belongie, Vladlen Koltun, and René Ranftl

  16. [24]

    Timo Lüddecke and Alexander Ecker. 2022. Image segmentation using text and image prompts. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 7086–7096

    2022

  17. [25]

    Muhammad Maaz, Hanoona Rasheed, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwer, and Ming-Hsuan Yang. 2022. Class-agnostic object detection with multi-modal transformer. InEuropean conference on computer vision. Springer, 512–531

    2022

  18. [26]

    Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi

  19. [27]

    Muhammad Arslan Manzoor, Sarah Albarri, Ziting Xian, Zaiqiao Meng, Preslav Nakov, and Shangsong Liang. 2023. Multimodality representation learning: A survey on evolution, pretraining and its applications.ACM Transactions on Multimedia Computing, Communications and Applications...

    2023

  20. [28]

    Maria-Elena Nilsback and Andrew Zisserman. 2008. Automated flower classifica- tion over a large number of classes. In2008 Sixth Indian conference on computer vision, graphics & image processing. IEEE, 722–729

    2008

  21. [29]

    Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. 2012. Cats and dogs. In2012 IEEE conference on computer vision and pattern recognition. IEEE, 3498–3505

    2012

  22. [30]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. InInternational conference on machine learnin...

    2021

  23. [31]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners.OpenAI blog 1, 8 (2019), 9

    2019

  24. [32]

    Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. 2022. Denseclip: Language-guided dense predic- tion with context-aware prompting. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 18082–18091

    2022

  25. [33]

    Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. 2019. Do imagenet classifiers generalize to imagenet?. InInternational conference on machine learning. PMLR, 5389–5400

    2019

  26. [34]

    Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. 2012. UCF101: A dataset of 101 human actions classes from videos in the wild.arXiv preprint arXiv:1212.0402(2012)

    2012 arXiv

  27. [35]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need.Advances in neural information processing systems30 (2017)

    2017

  28. [36]

    Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. 2019. Learning robust global representations by penalizing local predictive power.Advances in neural information processing systems32 (2019)

    2019

  29. [37]

    Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba

  30. [38]

    Hantao Yao, Rui Zhang, and Changsheng Xu. 2023. Visual-language prompt tun- ing with knowledge-guided context optimization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 6757–6767. Conference’17, July 2017, Washington, DC, USA Qiqi Zhan, ...

    2023

  31. [39]

    Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiao- dan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. 2021. Filip: Fine-grained interactive language-image pre-training.arXiv preprint arXiv:2111.07783(2021)

    2021 arXiv

  32. [40]

    Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. 2021. Florence: A new foundation model for computer vision.arXiv preprint arXiv:2111.11432 (2021)

    2021 arXiv

  33. [41]

    Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. 2022. Open-vocabulary detr with conditional matching. InEuropean conference on computer vision. Springer, 106–122

    2022

  34. [42]

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. 2023. Sig- moid loss for language image pre-training. InProceedings of the IEEE/CVF inter- national conference on computer vision. 11975–11986

    2023

  35. [43]

    Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexan- der Kolesnikov, and Lucas Beyer. 2022. Lit: Zero-shot transfer with locked-image text tuning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 18123–18133

    2022

  36. [44]

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. 2022. Conditional prompt learning for vision-language models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 16816–16825

    2022

  37. [45]

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. 2022. Learning to prompt for vision-language models.International Journal of Computer Vision 130, 9 (2022), 2337–2348

    2022

  38. [46]

    Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. 2023. Prompt- aligned gradient for prompt tuning. InProceedings of the IEEE/CVF international conference on computer vision. 15659–15669. A More Implementation details In this section, we provide further implementat...

    2023

  39. [2010]

    In2010 IEEE computer society conference on computer vision and pattern recognition

    Sun database: Large-scale scene recognition from abbey to zoo. In2010 IEEE computer society conference on computer vision and pattern recognition. IEEE, 3485–3492

  40. [2013]

    Fine-grained visual classification of aircraft.arXiv preprint arXiv:1306.5151 (2013)

    2013 arXiv

  41. [2021]

    InProceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Natural adversarial examples. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 15262–15271

  42. [2022]

    InThe Tenth International Con- ference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022

    Language-driven Semantic Segmentation. InThe Tenth International Con- ference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net. https://openreview.net/forum?id=RriDjddCLN

    2022

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.