A Paragraph is Worth a Thousand Captions: Rethinking Text Supervision for Vision-Language Retrieval

Amin Tabrizian; Arsyi Aziz; Binshuai Wang; Mahyar Ghazanfari; Peng Wei

REVIEW 2 major objections 6 minor 25 references

A Paragraph is Worth a Thousand Captions: Rethinking Text Supervision for Vision-Language Retrieval

T0 review · 2 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Paragraph supervision, not a longer context window, may be the real key to long-description retrieval.

desk verdict The DOCCI result is real and useful; the ShareGPT4V 'match' is likely inflated by train/eval overlap, so the headline claim needs qualification before it can be trusted. read the letter →

arxiv 2608.05260 v1 pith:45344GNC submitted 2026-08-05 cs.CV

Mahyar Ghazanfari , Amin Tabrizian , Arsyi Aziz , Binshuai Wang , Peng Wei This is my paper · ORCID

classification cs.CV

keywords vision-languageretrievaltextsupervisiongranularityparagraphdescriptionscontrastivelearningBLIPencoderfine-tuninglong-descriptionhardnegativestruncationanalysis

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that the granularity of the text used to train a vision-language model, rather than its architectural context length, is the main lever for retrieving images from long, detailed descriptions. By fine-tuning only the text encoder of a frozen BLIP model on either single captions or multi-sentence paragraphs, the authors isolate text granularity as the sole variable. Their paragraph-trained model matches Long-CLIP-L on ShareGPT4V and outperforms it by over 14 points on DOCCI image-to-text retrieval, without any positional-embedding or architecture changes. A sympathetic reader would care because it suggests a cheaper, data-centric path to long-text retrieval that competes with architectural extensions.

What carries the argument

The central object is the text-supervision granularity, varied across ten configurations from a single caption to a multi-sentence paragraph. The mechanism is a controlled training setup that freezes the BLIP vision encoder and fine-tunes only the text encoder and projection, so any performance difference is attributable to the text input. The paper also relies on a synthetic data pipeline using Qwen2-VL to generate diverse captions, hard negatives, and quality-scored paragraphs (with Llama 3.2 Vision as a scorer), and on a truncation analysis that shows how recall changes with token budget.

What would settle it

Re-run the ShareGPT4V evaluation excluding the 30K LLaVA-Pretrain images that overlap with CC3M (or use a strictly disjoint held-out set), and check whether the paragraph-trained model still matches Long-CLIP-L. Additionally, retrain the C5 configuration with paragraphs truncated to 60 tokens and verify whether the DOCCI and ShareGPT4V gains disappear, which would confirm that the benefit comes from processing the full paragraph rather than from the training data distribution alone.

Watch

Extended reading notes

Core claim

The paper's central claim is that training text granularity alone determines long-text retrieval performance, and that paragraph-level supervision is sufficient to match or beat an architecture built for long contexts. Holding the vision encoder fixed and fine-tuning only the BLIP text encoder, the authors find that a configuration using the original caption plus one generated paragraph (C5) reaches 66.1% I2T R@1 on DOCCI versus 51.6% for Long-CLIP-L, and matches Long-CLIP-L on ShareGPT4V (around 52% vs. 52.6%). They also report a mechanistic distinction: paragraph-trained models improve monotonically as token budget grows from 20 to 128, whereas caption-only training plateaus or degrades beyond 60 tokens, indicating that paragraph supervision teaches the text encoder to use extended sequences productively.

Load-bearing premise

The ShareGPT4V evaluation sample is assumed to be unseen during training, but 30K of its images come from LLaVA-Pretrain, which is built from CC3M, the same source as the paper's 500K training set—so those images may have been seen with other captions during fine-tuning.

Editorial extensions

If this is right

If paragraph supervision is the main driver, practitioners can improve long-description retrieval by investing in richer text annotations rather than extending context windows or modifying positional embeddings.
Caption-only fine-tuning on synthetic captions may actively degrade long-description retrieval, so data curation for retrieval should prioritize paragraph-level descriptions.
Hard negative captions, typically beneficial in contrastive learning, are harmful in a text-only fine-tuning setting with a frozen vision encoder, suggesting that negative sampling strategies need to be re-evaluated for partial-update training.
The monotonic improvement of paragraph-trained models with token budget implies that longer paragraphs (beyond the 128-token limit) could unlock further gains, and that the text encoder's capacity is not the bottleneck.
The Pareto analysis shows that paragraph-containing configurations achieve large long-description gains with only modest short-caption trade-offs, offering a practical operating point for retrieval systems that must handle both regimes.

Reading between the lines

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

The paper's finding suggests a data-centric alternative to architectural fixes: even a modest amount of paragraph data (25% of 500K images) captures most of the long-text benefit, implying that collecting or generating paragraph annotations could be a cost-effective upgrade for existing CLIP-style models.
The degradation of caption-only training on long descriptions hints that the text encoder, when fine-tuned on short texts, actively narrows its representational capacity; this could be tested by evaluating such models on zero-shot long-text tasks beyond retrieval, such as image captioning with detailed prompts.
The success of paragraphs over multiple captions, despite the latter providing more textual diversity, suggests that coherence and structure matter more than raw token count; this could be probed by ablating paragraph structure (e.g., scrambling sentences) while keeping the same vocabulary.
The hard-negative collapse in the text-only fine-tuning regime may be a specific instance of a broader phenomenon: when one encoder is frozen, negative sampling that relies on joint adaptation can push the trainable encoder into a degenerate subspace; this hypothesis could be tested by unfreezing the vision encoder for a few epochs and observing whether the hard-negative damage reverses.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The paper is worth a look, and the referee time is justified, but the headline claim needs a carve-out. The DOCCI result is the real finding: a BLIP text encoder fine-tuned on paragraphs, with the vision encoder frozen, beats Long-CLIP-L by 14 points on image-to-text R@1 on a benchmark none of the models were trained on. That is a large, believable effect, and it is backed by a clean controlled design: ten text-supervision configurations, a shared frozen vision encoder, and a synthetic paragraph pipeline with a quality-scoring loop. The truncation analysis is the most instructive piece—paragraph-trained models keep gaining recall as the token budget grows to 128, while caption-trained models plateau or drop past 60 tokens. That is a genuine mechanistic observation.

The soft spots are real but fixable. The most serious is the ShareGPT4V evaluation: the 80K sample includes 30K LLaVA-Pretrain images, which are CC3M images, and the training set is 500K CC3M images. The paper reports no overlap filtering. A random 500K sample of CC3M would overlap a random 30K LLaVA-Pretrain sample by thousands of images (roughly 15% of that subset). At evaluation, the model has already been fine-tuned to align those exact images with long paragraph text, so the "matches Long-CLIP-L on ShareGPT4V without having seen the data" claim is inflated by an unknown amount. The gap to Long-CLIP-L is 0.6-0.9 points; even partial overlap could erase it. This is not a fatal flaw because the DOCCI result is unaffected, but the abstract should be toned down.

Two smaller issues. The paper uses a nonstandard full-set retrieval protocol on Flickr30k and COCO but does not state explicitly that the Long-CLIP and CLIP baselines were recomputed under the same protocol. If those numbers come from the original papers, the comparison is invalid. And there are no error bars anywhere; the ShareGPT4V difference is within the range where seed-to-seed variance matters.

The hard-negative finding—collapse from ~50% to ~25% with a single hard negative—is surprising and well documented; I'd want to see it replicated under a second seed before treating it as robust, but the direction is plausible given the frozen vision encoder can't co-adapt. The data-efficiency result that 25% of paragraphs does best on short-caption retrieval is similarly provisional.

No circularity or invented entities; the claims are empirical comparisons against external benchmarks. The paper is honest about its scope.

For a referee: send it out, but require an overlap analysis for ShareGPT4V, an explicit statement that baselines were rerun on the same splits, and error bars on the headline numbers. The authors have a solid contribution underneath; it deserves a proper review rather than a desk rejection.

Referee Report

2 major / 6 minor

Summary. The paper studies how the granularity of text supervision affects contrastive vision-language retrieval. The authors construct a synthetic dataset from 500K CC3M images using Qwen2-VL to generate five positive captions, five hard-negative captions, and one multi-sentence paragraph per image, with Llama 3.2 Vision scoring paragraph faithfulness and triggering regeneration below score 5. They fine-tune only the BLIP text encoder and text projection, freezing the vision encoder, across 10 configurations that vary caption count, paragraph presence, and hard-negative inclusion. Evaluations are on Flickr30k, COCO (full-set), an 80K ShareGPT4V sample, and DOCCI. The main claims are that paragraph-supervised models match Long-CLIP-L on ShareGPT4V and exceed it by more than 14 points on DOCCI image-to-text R@1 without architectural changes; caption-only fine-tuning degrades long-description retrieval; more caption diversity helps short captions but hurts long descriptions; hard negatives are harmful under text-only fine-tuning; and truncation analysis shows paragraph-trained models use longer token budgets effectively.

Significance. If the results hold, the paper provides a clean empirical dissociation of text granularity from architectural and data-scale confounds. The frozen-vision design across 10 configurations is a useful experimental contribution, and the DOCCI benchmark is genuinely unseen by all training stages, so the >14-point I2T gain over Long-CLIP-L is a substantial result. The truncation analysis offers a plausible mechanism rather than a black-box comparison. The paper contains no fitted constants and its evidence is empirical comparisons against external benchmarks, which is a strength. The main weakness is the ShareGPT4V evaluation: because the 80K sample includes 30K LLaVA-Pretrain images from the CC3M source used for training, the 'unseen' claim for that benchmark is not currently established. The central contribution may survive on DOCCI alone, but the headline claim needs to be repaired with an overlap analysis.

major comments (2)

[Sec. 4.2, Table 2] The claim that paragraph-supervised models 'match Long-CLIP-L on ShareGPT4V ... without having seen any ShareGPT4V data' is not supported as stated because the evaluation set overlaps with the training data source. The 80K ShareGPT4V sample is described as 50K COCO images plus 30K LLaVA-Pretrain images, and LLaVA-Pretrain is built from CC3M, the same dataset from which the 500K training images are sampled in Sec. 3.1. If those 30K images are present in the 500K training set, the text encoder was fine-tuned on synthetic paragraphs for exactly those evaluation images, which can inflate the I2T R@1 reported for C4-C6. No overlap filtering or overlap statistics are reported anywhere, and the Limitations section (Sec. 6) does not mention this risk. Please compute the overlap by image ID or perceptual hash, report results on the overlapping and non-overlapping subsets separately, and either remove or carefully qualify the 'unseen' claim. The DOCCI results are on a distinct dataset and are not affected by this issue.
[Sec. 3.4, Table 1] The full-set evaluation protocol is a deliberate departure from the standard 1K/5K Flickr30k and COCO splits, but the manuscript does not state whether all baselines (BLIP0, CLIP0, Long-CLIP-B, Long-CLIP-L) were evaluated by the authors with identical preprocessing, tokenization, and retrieval code. Because full-set numbers are not directly available from prior publications, the comparisons cannot be verified without this information. In addition, Long-CLIP is described as fine-tuned on 1M ShareGPT4V pairs (Sec. 4.2), and the ShareGPT4V evaluation sample in Table 2 includes 50K COCO images; evaluating Long-CLIP on the full COCO train2017 set may therefore be favorable to Long-CLIP because of training-data overlap. Please state explicitly how each baseline was obtained, add any overlap filtering, and consider reporting standard-split results in a supplementary table for comparability with the literature.

minor comments (6)

[Sec. 4.2 vs. Table 4] The token-budget treatment of ShareGPT4V differs between Table 2 (77 tokens) and Table 4 (128 tokens for the same C5 checkpoint), yielding 51.7% versus 63.1% I2T R@1. The protocol note in Table 4 is important and should be stated in Sec. 4.2 as well, so that the reader understands that the 'match' claim is a conservative, 77-token comparison and that the model improves substantially at its native training budget.
[Sec. 5.1, Table 3] The paper should specify the exact positive and negative counts for the 'Negative' rows in Table 3; as written, the reader cannot tell how many positive captions accompany the 1, 3, or 5 hard negatives.
[Sec. 5.2, Table 4] The column header 'CFG5' appears to be a typo for C5; please correct it.
[Sec. 6] The hard-negative finding is stated as if it applies to hard negatives generally, but the experiments use synthetic factually-incorrect captions in the denominator with a frozen vision encoder; the conclusions should be scoped to this setup, since standard in-batch hard negatives may behave differently.
[Abstract and Sec. 3.4] The statement 'Code and dataset are available here' does not include a URL; please provide a working link or repository identifier.
[Sec. 4.3, Fig. 2] The truncation analysis uses 10K samples from ShareGPT4V and DOCCI without specifying the selection method or random seed; please add this information for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all claims are empirical comparisons against external benchmarks, with no fitted parameters renamed as predictions and no derivation that reduces to its own inputs.

full rationale

The paper is an empirical study, not a derivation. It fine-tunes a BLIP text encoder on synthetic captions/paragraphs generated offline for 500K CC3M images and compares the resulting checkpoints against external baselines (BLIP0, CLIP0, Long-CLIP-B/L) on Flickr30k, COCO, ShareGPT4V, and DOCCI. The central claim, that paragraph supervision matches or exceeds Long-CLIP on long-description retrieval, is supported by reported R@1 numbers on benchmarks that were not used for training. There are no fitted constants whose values are subsequently reported as predictions: the configurations differ only in text inputs and loss formulations, and the same hyperparameters are fixed across all runs. The paragraph data were generated before training by an LLM pipeline with quality scoring, not selected post hoc to match evaluation labels. Cited prior work is used for context and baselines, not as a substitute for the paper's own evidence, and no load-bearing argument rests on a self-citation. The one substantive concern identified in the manuscript context is that the fixed 80K ShareGPT4V evaluation sample contains 30K LLaVA-Pretrain images, which derive from CC3M, the same source as the 500K training set; if those images overlap with training, the ShareGPT4V result could be inflated. That is a data-leakage and benchmark-validity concern, not a circularity concern: it does not make any 'prediction' equivalent to an input by construction. The DOCCI results, obtained on a dataset unseen by all models, are independent evidence for the paper's main direction. Therefore the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claims rest on the synthetic data pipeline, the frozen-vision experimental design, and external benchmarks. No physical or theoretical entities are introduced. The main ledger items are hand-chosen pipeline thresholds and unverified assumptions about data quality and train/eval disjointness.

free parameters (3)

Paragraph quality regeneration threshold = Llama score > 5, up to 3 attempts
Hand-chosen threshold in Sec. 3.1; determines which paragraphs are kept and therefore the quality of paragraph supervision.
Token budgets = 77 for captions, 128 for paragraphs
Design choice in Sec. 3.2 bounding text granularity; the 128-token paragraph limit is a stated limitation and affects long-description results.
Training hyperparameters = lr=1e-5, batch=256, epochs=10, tau=0.07
Fixed across all configurations (Sec. 3.2); standard values not fitted to benchmarks, but they shape every result.

assumptions (4)

domain assumption Generated captions and paragraphs are factually correct positive descriptions of their images.
Relies on Qwen2-VL generation; a 1,000-image GPT-4o and human audit (Sec. C) supports this, but errors in the remaining samples would inject label noise.
domain assumption Frozen BLIP vision encoder provides a stable image manifold, so all performance differences are attributable to the text encoder.
Core design premise of Sec. 3.2; authors acknowledge that text-only fine-tuning shifts the text manifold and degrades T2I relative to BLIP0.
domain assumption Hard negative captions are true negatives in the contrastive batch, not descriptions that match other images.
Sec. 3.1 generates negatives per image, but Sec. 5.1 uses them across the batch; a negative for one image may be a valid caption for another image, adding unintended label noise.
domain assumption The ShareGPT4V evaluation sample is disjoint from the 500K CC3M training images.
Table 2 note says the 80K sample contains 30K LLaVA-Pretrain images; LLaVA-Pretrain is built from CC3M, so this assumption is questionable and has not been checked in the paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Paragraph is Worth a Thousand Captions: Rethinking Text Supervision for Vision-Language Retrieval." pith.science (2026). https://pith.science/paper/45344GNC

@misc{pith2026260805260,
  author       = {Pith},
  title        = {Pith review of: A Paragraph is Worth a Thousand Captions: Rethinking Text Supervision for Vision-Language Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/45344GNC}},
  note         = {Machine review of arXiv:2608.05260}
}

read the original abstract

Contrastive vision-language models such as CLIP and BLIP are typically trained on short image captions, limiting their ability to retrieve images from detailed textual descriptions. While methods such as Long-CLIP extend the token limit through positional embedding interpolation, we ask a simpler question: does training text granularity alone determine long-text retrieval performance? We present a systematic study of supervision ranging from single captions to multi-sentence paragraphs for contrastive image-text retrieval. Using a synthetic pipeline based on Qwen2-VL and Llama 3.2 Vision, we generate diverse captions, hard negatives, and quality-scored paragraphs for 500K CC3M images. To isolate the effect of text granularity, we fine-tune only the BLIP text encoder while keeping the vision encoder frozen across 10 training configurations. Our paragraph-supervised models match Long-CLIP-L on ShareGPT4V and outperform it by more than 14 points on DOCCI for image-to-text retrieval, without architectural changes. We further show that paragraph supervision enables effective use of long token sequences, whereas caption-only training degrades beyond 60 tokens. Increasing caption diversity improves short-caption retrieval with diminishing returns, while paragraph supervision consistently benefits long-description benchmarks and hard negatives prove detrimental in text-only fine-tuning. Evaluations on Flickr30k, COCO, ShareGPT4V, and DOCCI provide a comprehensive analysis of the trade-offs between text granularity, retrieval direction, and description length.

Figures

Figures reproduced from arXiv: 2608.05260 by the authors.

**Figure 1.** Overview of our framework. Top: The synthetic data pipeline uses Qwen2-VL to generate captions, hard negatives, and paragraphs, with Llama 3.2 Vision scoring paragraphs in a feedback loop. Bottom: Training fine-tunes only the text encoder and text projection, with precomputed image embeddings from a frozen BLIP vision encoder. 256-dimensional image embeddings v = f img proj(Φ img(I)) for all training images, where Φ… view at source ↗

**Figure 2.** R@1 as a function of maximum token length during text encoding, evaluated on 10K samples each from (a) ShareGPT4V and (b) DOCCI. Paragraph-trained models (C4, C5, C9) show monotonically increasing recall with longer sequences on both datasets, while caption-trained models (C1, C3, C8) plateau around 60–80 tokens on ShareGPT4V and 40–60 tokens on DOCCI, after which performance stagnates or degrades [PITH_FULL_IMAG… view at source ↗

**Figure 3.** Pareto trade-off between short-caption (Flickr30k + COCO) and longdescription (ShareGPT4V + DOCCI) average R@1. Paragraph-trained configurations (C4–C6, C9) cluster in the upper region, achieving large long-description gains with modest short-caption trade-offs. C5 sits on the Pareto frontier in the I2T direction. Dashed orange line indicates the Pareto frontier. all paragraph configurations achieve similar T2I per… view at source ↗

Figures from the paper (14 more)

**Figure 4.** Figure 4: visualizes the distributions of paragraph lengths, original caption lengths, and Llama quality scores. Paragraphs follow a roughly Gaussian distribution centered around 90–95 words, while original captions are sharply concentrated below 15 words, highlighting the lar…

**Figure 5.** Figure 5: Dataset quality: GPT-4o (with image) vs. human annotation score distributions for positive captions, hard negative captions, and paragraphs. D Training Dynamics [PITH_FULL_IMAGE:figures/full_fig_p021_5.png]

**Figure 6.** Figure 6: Training dynamics across all configurations. (a) Training loss. (b) Validation loss. (c) Validation I→T R@1. (d) Validation T→I R@1. Paragraph-trained models converge to lower loss and higher R@1 than caption-only models. C7 (hard negatives, dashed) exhibits increasing…

**Figure 7.** Figure 7: Validation I→T R@1 convergence by configuration group. (a) Caption-only configurations show varied convergence between 85–98%. (b) Paragraph configurations converge rapidly to >99.8% by epoch 2 [PITH_FULL_IMAGE:figures/full_fig_p023_7.png]

**Figure 8.** Figure 8: t-SNE visualization of image (blue) and text (red) embeddings on Flickr30k. All configurations preserve a two-hemisphere layout; paired cosine similarities are tightly clustered around 0.43–0.50. preserved after text-only fine-tuning when the evaluation text is short. …

**Figure 9.** Figure 9: t-SNE visualization on COCO. Similar to Flickr30k, all configurations exhibit clean modality separation, reflecting that the benchmark descriptions are short and the gap in text-granularity between training and evaluation is small. for the catastrophic failure observed…

**Figure 10.** Figure 10: t-SNE visualization on ShareGPT4V. Paragraph-trained configurations (C4– C6, C9) produce visibly intermixed image and text distributions, with C9 (cos=0.658) approaching a unimodal cloud. Caption-only configurations (C1, C3, C8) and the hard-negative configuration C7 …

**Figure 11.** Figure 11: t-SNE visualization on DOCCI. The same pattern as on ShareGPT4V emerges on this unseen benchmark: paragraph configs (C4–C6, C9) produce substantially more mixed distributions and higher paired cosines (0.53–0.64) than caption-only configs (0.42–0.47), with C7 collapsi…

**Figure 12.** Figure 12: Mean paired image–text cosine similarity across the four benchmarks for nine representative configurations. Paragraph-trained configurations (C4–C6, C9) produce large alignment gains on ShareGPT4V and DOCCI while leaving Flickr30k and COCO alignment approximately unch…

**Figure 13.** Figure 13: Qualitative text-to-image retrieval on DOCCI (sample 1). Top-5 retrieved images per model. Only C5 (ours) retrieves the ground truth at rank 1. often sharing high-level scene category with the query but missing fine-grained attributes such as text on signage, material…

**Figure 14.** Figure 14: Qualitative text-to-image retrieval on DOCCI (sample 2). Top-5 retrieved images per model. Only C5 (ours) retrieves the ground truth at rank 1. is a differentiable function of I. We compute the per-pixel saliency as the absolute input gradient, reduced across the thre…

**Figure 15.** Figure 15: Qualitative text-to-image retrieval on DOCCI (sample 3). Top-5 retrieved images per model. Only C5 (ours) retrieves the ground truth at rank 1. broad portions of the image rather than collapsing to a single point, and tend to highlight regions that correspond to salie…

**Figure 16.** Figure 16: Qualitative text-to-image retrieval on DOCCI (sample 4). Top-5 retrieved images per model. Only C5 (ours) retrieves the ground truth at rank 1 [PITH_FULL_IMAGE:figures/full_fig_p032_16.png]

**Figure 17.** Figure 17: Gradient-based saliency maps on DOCCI. For each sample, the full description query is shown above, followed by the original image and heatmap overlays for five models. Warm regions (red/yellow) indicate pixels that most influence the text–image similarity score [PITH…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 17 canonical work pages

[1]

In: Eu- ropean Conference on Computer Vision (ECCV)

Chen, L., Li, J., Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: ShareGPT4V: Improving large multi-modal models with better captions. In: Eu- ropean Conference on Computer Vision (ECCV). pp. 370–387. Springer (2024)

work page 2024
[2]

Advances in Neural Information Processing Systems (NeurIPS) 36, 35544–35575 (2023)

Fan,L.,Krishnan, D.,Isola,P., Katabi,D.,Tian,Y.:ImprovingCLIPtrainingwith language rewrites. Advances in Neural Information Processing Systems (NeurIPS) 36, 35544–35575 (2023)

work page 2023
[3]

arXiv preprint arXiv:2407.21783 (2024)

Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Let- man,A., Mathur, A.,Schelten,A., Vaughan, A.,etal.: TheLlama 3herd of models. arXiv preprint arXiv:2407.21783 (2024)

arXiv 2024
[4]

In: International Conference on Machine Learning (ICML)

Jia,C.,Yang,Y.,Xia,Y., Chen,Y.T.,Parekh,Z., Pham,H.,Le,Q.,Sung,Y.H.,Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning (ICML). pp. 4904–4916. PMLR (2021)

work page 2021
[5]

In: European Conference on Computer Vision (ECCV)

Lai, Z., Zhang, H., Zhang, B., Wu, W., Bai, H., Timofeev, A., Du, X., Gan, Z., Shan, J., Chuah, C.N., et al.: VeCLIP: Improving clip training via visual-enriched captions. In: European Conference on Computer Vision (ECCV). pp. 111–127. Springer (2024)

work page 2024
[6]

In: Proceedings of the International Conference on Machine Learning (ICML) (2024)

Lavoie, S., Kirichenko, P., Ibrahim, M., Assran, M., Wilson, A.G., Courville, A., Ballas, N.: Modeling caption diversity in contrastive vision-language pretraining. In: Proceedings of the International Conference on Machine Learning (ICML) (2024)

work page 2024
[7]

In: International Con- ference 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 Con- ference on Machine Learning (ICML). pp. 12888–12900. PMLR (2022)

work page 2022
[8]

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)

work page 2014

Show all 25 references

[9]

In: International Conference on Learning Representations (ICLR) (2019)

Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (ICLR) (2019)

2019
[10]

In: ECCV (2024)

Onoe, Y., Rane, S., Berger, Z., Bitton, Y., Cho, J., Garg, R., Ku, A., Parekh, Z., Pont-Tuset, J., Tanzer, G., Wang, S., Baldridge, J.: DOCCI: Descriptions of Connected and Contrasting Images. In: ECCV (2024)

2024
[11]

arXiv preprint arXiv:1807.03748 (2018)

Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predic- tive coding. arXiv preprint arXiv:1807.03748 (2018)

2018 arXiv
[12]

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

Radenovic, F., Dubey, A., Kadian, A., Mihaylov, T., Vandenhende, S., Patel, Y., Wen, Y., Ramanathan, V., Mahajan, D.: Filtering, distillation, and hard negatives for vision-language pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni...

2023
[13]

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
[14]

arXiv preprint arXiv:2204.06125 1(2), 3 (2022)

Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text- conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)

2022 arXiv
[15]

In: International Conference on Learning Representations (ICLR) (2021) 16 M.Ghazanfari et al

Robinson, J.D., Chuang, C.Y., Sra, S., Jegelka, S.: Contrastive learning with hard negative samples. In: International Conference on Learning Representations (ICLR) (2021) 16 M.Ghazanfari et al

2021
[16]

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 (2022)

2022
[17]

In: Proceedings of the 3rd Workshop on Advances in Language and Vision Research (ALVR)

Rösch, P.J., Oswald, N., Geierhos, M., Libovický, J.: Enhancing conceptual un- derstanding in multimodal contrastive learning through hard negative samples. In: Proceedings of the 3rd Workshop on Advances in Language and Vision Research (ALVR). pp. 102–115. Association for Com...

2024 doi
[18]

In: Proceed- ings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Proceed- ings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2556–25...

2018
[19]

In: International Con- ference on Learning Representations (ICLR) Workshop (2014)

Simonyan, K., Vedaldi, A., Zisserman, A.: Deep inside convolutional networks: Visualising image classification models and saliency maps. In: International Con- ference on Learning Representations (ICLR) Workshop (2014)

2014
[20]

In: 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Wang, H., Ju, C., Lin, W., Xiao, S., Chen, M., Huang, Y., Liu, C., Yao, M., Lan, J., Chen, Y., Liu, Q., Wang, Y.: Advancing myopia to holism: Fully contrastive language-image pre-training. In: 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2979...

2025
[21]

arXiv preprint arXiv:2409.12191 (2024)

Wang, P., Bai, S., Tan, S., Wang, S., Fan, Z., Bai, J., Chen, K., Liu, X., Wang, J., Ge, W., et al.: Qwen2-VL: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191 (2024)

2024 arXiv
[22]

Transactions of the Association for Computational Linguistics (ACL)2, 67–78 (2014)

Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics (ACL)2, 67–78 (2014)

2014
[23]

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

Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid loss for language im- age pre-training. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 11975–11986 (2023)

2023
[24]

In: European Conference on Computer Vision (ECCV)

Zhang, B., Zhang, P., Dong, X., Zang, Y., Wang, J.: Long-CLIP: Unlocking the long-text capability of CLIP. In: European Conference on Computer Vision (ECCV). pp. 310–325. Springer (2024)

2024
[25]

TRUTH / IS A BREATH AWAY / THE VICTORY

Zheng, K., Zhang, Y., Wu, W., Lu, F., Ma, S., Jin, X., Chen, W., Shen, Y.: Dream- LIP: Language-image pre-training with long captions. In: European Conference on Computer Vision (ECCV) (2024) A Paragraph is Worth a Thousand Captions 17 Supplementary Material A Paragraph is Wor...

2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

In: Eu- ropean Conference on Computer Vision (ECCV)

Chen, L., Li, J., Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: ShareGPT4V: Improving large multi-modal models with better captions. In: Eu- ropean Conference on Computer Vision (ECCV). pp. 370–387. Springer (2024)

work page 2024

[2] [2]

Advances in Neural Information Processing Systems (NeurIPS) 36, 35544–35575 (2023)

Fan,L.,Krishnan, D.,Isola,P., Katabi,D.,Tian,Y.:ImprovingCLIPtrainingwith language rewrites. Advances in Neural Information Processing Systems (NeurIPS) 36, 35544–35575 (2023)

work page 2023

[3] [3]

arXiv preprint arXiv:2407.21783 (2024)

Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Let- man,A., Mathur, A.,Schelten,A., Vaughan, A.,etal.: TheLlama 3herd of models. arXiv preprint arXiv:2407.21783 (2024)

arXiv 2024

[4] [4]

In: International Conference on Machine Learning (ICML)

Jia,C.,Yang,Y.,Xia,Y., Chen,Y.T.,Parekh,Z., Pham,H.,Le,Q.,Sung,Y.H.,Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning (ICML). pp. 4904–4916. PMLR (2021)

work page 2021

[5] [5]

In: European Conference on Computer Vision (ECCV)

Lai, Z., Zhang, H., Zhang, B., Wu, W., Bai, H., Timofeev, A., Du, X., Gan, Z., Shan, J., Chuah, C.N., et al.: VeCLIP: Improving clip training via visual-enriched captions. In: European Conference on Computer Vision (ECCV). pp. 111–127. Springer (2024)

work page 2024

[6] [6]

In: Proceedings of the International Conference on Machine Learning (ICML) (2024)

Lavoie, S., Kirichenko, P., Ibrahim, M., Assran, M., Wilson, A.G., Courville, A., Ballas, N.: Modeling caption diversity in contrastive vision-language pretraining. In: Proceedings of the International Conference on Machine Learning (ICML) (2024)

work page 2024

[7] [7]

In: International Con- ference 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 Con- ference on Machine Learning (ICML). pp. 12888–12900. PMLR (2022)

work page 2022

[8] [8]

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)

work page 2014