Pith. sign in

REVIEW 5 major objections 5 minor 32 references

The paper claims that text-to-video retrieval improves when alignment happens between Gaussian distributions rather than single points, using a deterministic diffusion-inspired bridge and a KL-based contrastive loss, and that the resulting

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 08:46 UTC pith:ALO456CD

load-bearing objection Real extension of DITS to distribution space, but the inference is late-interaction while baselines are bi-encoders, so the +4 R@1 claim needs a controlled comparison. the 5 major comments →

arxiv 2607.20984 v1 pith:ALO456CD submitted 2026-07-23 cs.CV

Distribution-Alignment Bridge for Uncertainty-Aware Text-to-Video Retrieval

Kyeongmo Chae , Jihoon Lee , Sangtae Ahn This is my paper
classification cs.CV
keywords text-to-video retrievalprobabilistic embeddingdistribution alignmentuncertainty modelingdiffusion-inspired refinementcontrastive learningKL divergenceCLIP
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that text-to-video retrieval improves when the matching is done between probability distributions rather than single embedding points. It encodes each text and video as a Gaussian (mean for semantics, variance for uncertainty), then runs a deterministic diffusion-inspired bridge that iteratively refines the text distribution toward the video distribution over truncated steps. A bidirectional contrastive loss over directional KL divergences makes ranking sensitive to both the means and the variances. On MSR-VTT the method reaches 56.2 R@1 versus 51.9 for the point-wise DITS baseline and cuts mean rank from 11.6 to 4.1; it also reports gains on MSVD and VATEX. The same bridge-induced margin behaves as a calibrated confidence signal, giving 85.6% R@1 when only the 25% most confident queries are answered.

Core claim

The central claim is that deterministic distribution-space refinement yields a better text-video embedding space than point-wise refinement, and that the induced distributional margins are a usable uncertainty signal. DAB maps CLIP features to Gaussian parameters (µ, log σ²), with a token-based branch estimating video variance from frame variability. The Distribution Bridge then predicts drift offsets (∆µ, ∆log σ²) over truncated timesteps, updating the text distribution toward the video distribution without stochastic sampling. Training uses a bidirectional contrastive loss over directional KL divergences, so matched pairs need lower KL cost than in-batch negatives but are not forced to sha

What carries the argument

The Distribution Bridge is the central mechanism: a deterministic, diffusion-inspired refinement that operates on (µ, log σ²) instead of on feature vectors. A time-conditioned drift MLP predicts mean and log-variance offsets, scaled by a step-size schedule derived from a linear beta schedule; updates subtract the scaled offsets over T′=32 truncated steps, with log-variance clipped to [−6,2]. Zero-initialized drift heads and a contraction intuition (each step shrinks the residual by at least a factor 1−χt, with χt ∈ [0.2,0.6]) keep the trajectory stable. The probabilistic head with a token-based video variance branch supplies the distributions the bridge transforms, and the PAC loss with dire

Load-bearing premise

That the reported gains come from the distribution-alignment mechanism itself rather than from DAB's more generous training protocol—the component ablation shows the bridge alone can lower R@1 against the X-Pool baseline, so the protocol explains part of the improvement.

What would settle it

Run the strongest point-wise baseline under DAB's exact training protocol (10 epochs, batch size 32, 12 sampled frames, same CLIP backbone and chunk sizes); if its R@1 reaches DAB's 56.2, the gain is protocol, not the bridge. Alternatively, check whether the calibration curve survives when queries are sorted by a score margin computed without distributions; if a random or score-based margin gives the same 85.6% R@1 at 25% coverage, the uncertainty-calibration claim is not specific to the bridge.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Retrieval can use a deterministic, sampling-free diffusion-style refinement, avoiding the instability and variance of generative diffusion while still iterating toward the target distribution.
  • Ranking costs that account for variance (KL) outperform ones that only penalize mean geometry (W2), changing the choice of divergence in contrastive retrieval.
  • The model's learned uncertainty is preserved, not collapsed: video variance stays several times larger than text variance, so refinement does not force the modalities to share a variance scale.
  • The distributional margin between top-ranked candidates is a usable confidence score, enabling selective prediction where the system answers only high-confidence queries.

Where Pith is reading between the lines

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

  • Because the same CLIP features and cross-attention pooling are shared across many retrieval architectures, the distribution bridge could be grafted onto other CLIP-based text-video models, but whether the gains transfer is untested.
  • The calibration result suggests a lightweight confidence estimator for any retrieval system that exposes top-2 scores, independent of distributions, but the paper does not compare against that cheaper baseline.
  • The component ablation (bridge alone lowers R@1 relative to the X-Pool baseline) implies the bridge's benefit may depend on training it together with the probabilistic head; a fair test would hold the full protocol fixed and ablate only the bridge.
  • One natural extension is to use the predicted variance to modulate retrieval in high-ambiguity settings, such as query rewriting or human confirmation, though the paper does not demonstrate this.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. DAB recasts text-to-video retrieval as distribution-to-distribution alignment: text and video embeddings are mapped to Gaussian parameters (mean and log-variance), and a deterministic, diffusion-inspired bridge iteratively refines the text distribution toward a target video distribution over truncated steps. Training uses a bidirectional contrastive loss over directional KL costs between the refined text distribution and video distributions. The paper reports strong gains over published baselines on MSR-VTT, MSVD, and VATEX (e.g., MSR-VTT R@1 56.2 vs. 51.9 for DITS, MnR 4.1 vs. 11.6), together with component ablations and diagnostic analyses of uncertainty calibration and semantic-neighborhood organization.

Significance. If the empirical claims hold, the paper would provide an interesting deterministic alternative to stochastic diffusion-based retrieval and a way to incorporate uncertainty into contrastive ranking. The architecture is coherent and the paper is clearly written. Strengths include the explicit distribution-space formulation, the closed-form KL cost, the transparent reporting of hyperparameters, and the inclusion of component ablations and uncertainty diagnostics. However, the headline results rest on two premises that are not currently established: that the comparison to published baselines is fair under the same training/inference protocol, and that the observed gains come from distribution-space refinement rather than from the architecture's per-pair late-interaction inference. The stability argument in Section 3.3 is explicitly conditional on an unproven assumption. These issues are load-bearing for the central empirical claim, so a revision is needed before the paper can be accepted.

major comments (5)
  1. [§3.3, §4.1 (Eqs. 7, 10)] The inference procedure is underspecified and creates an apples-to-oranges comparison. The drift network in Eq. (7) takes the target video distribution (μ_v, σ_v) as input, and Eq. (10) defines C_ij using the bridged text distribution P_hat_ti. The paper never states what target is used at inference; the only coherent reading is that each query text distribution is refined separately against every gallery video. This makes DAB a per-pair late-interaction model. Most compared baselines — X-Pool, UATVR, DITS, T-MASS — are bi-encoders that score with a fixed query embedding. The +4.3 R@1 / 64% MnR improvement over DITS could be substantially attributable to this interaction asymmetry rather than to distribution-space refinement per se. Please state the inference protocol explicitly and include a controlled comparison: either compare against a late-interaction version of a bi-encoder baselin
  2. [§4.2, Tables 1–3] All baseline numbers are quoted from original papers rather than re-run under DAB's training protocol (10 epochs, batch size 32, 12 sampled frames, dropout 0.3/0.4, inference chunk sizes 128/512). Because retrieval results are highly sensitive to training budget and implementation details, the reported improvements over DITS, UATVR, and others may reflect a different budget rather than the proposed distribution bridge. Please re-run the key baselines (at least X-Pool, DITS, and UATVR) in the same codebase and under the same epoch/batch/frame settings, or otherwise demonstrate that the published numbers are directly comparable.
  3. [Table 5 (Component ablation)] The component ablation does not isolate the contribution of the bridge. Adding the bridge to the X-Pool baseline lowers R@1 from 46.9 to 43.2; adding the probabilistic head alone gives 44.3; only the full combination reaches 52.6. The text describes the bridge as 'most effective only when guided by uncertainty,' but no variant shows the incremental benefit of the bridge while holding the probabilistic head fixed (e.g., 'w/ Prob.' vs. 'w/ Prob. + Bridge'). The current ablation leaves open the possibility that the probabilistic embedding alone, or an interaction between components, drives the gains. Please report the incremental bridge effect over the probabilistic head, and likewise the incremental probabilistic-head effect over the bridge.
  4. [§3.3 (stability argument)] The stability argument is conditional on an unproven assumption: 'If the PAC-trained drift approximates the residual x_t − x*…'. No evidence is provided that the drift network approximates this residual; the PAC loss is a contrastive objective over distributional costs, not a regression loss to x_t − x*. As written, the contraction claim (rate at most 0.8) is therefore not established. Please either prove or empirically verify the residual approximation, or weaken the claim to a heuristic observation. This does not invalidate the method, but it is a stated theoretical justification and should not remain unsupported.
  5. [§4.4] The 'calibrated uncertainty-aware ranking' claim is stronger than the evidence. The paper shows that sorting queries by the bridge-induced KL margin improves selective R@1 (85.6% at 25% coverage vs. 56.2% random), but it does not compare this confidence signal against simpler alternatives such as text-variance magnitude or the raw score margin from a bi-encoder baseline. It also uses the term 'calibrated' without reporting calibration metrics (e.g., ECE or reliability diagrams). Please either add such comparisons, or soften the claim to 'selective prediction improves when ranking by KL margin.'
minor comments (5)
  1. [§3.3, Eq. (7)] The notation uses σ_t, σ_v for inputs to the drift network, but the probabilistic head outputs log σ²; please clarify whether the network receives σ or log σ², and be consistent with the update equations.
  2. [§4.4, Fig. 4] AURC is reported without a definition. Please define the risk–coverage curve and how the area is computed; also add error bars or multiple seeds, since all experiments use a single seed (seed 24).
  3. [§4.1] The seed for all runs is only one value (seed 24), and no standard deviations are reported. For a paper whose central claim is empirical superiority, a few repeat runs (or at least results over 2–3 seeds) would substantially increase confidence.
  4. [§3.4] The acronym PAC is used for 'Probabilistic Alignment Contrastive' and later in §3.3 for 'PAC-trained drift.' This dual use is confusing; consider renaming the trained drift to avoid ambiguity.
  5. [§4.3, Table 4] In the caption and text, 'T ′=8' and 'T ′=64' rows are reported as 'T ′' but the notation is not defined at first use in the ablation section; a short definition would help.

Circularity Check

0 steps flagged

No significant circularity; the paper's claims are empirical and its derivation chain does not reduce to its inputs.

full rationale

The DAB framework is an empirical retrieval model; its headline results are benchmark measurements, not derivations from fitted quantities. The distribution bridge (Eqs. 7-9) is trained with a contrastive KL loss (Eq. 12), and the component ablation in Table 5 explicitly shows that the bridge alone lowers R@1 versus the X-Pool baseline, which is the opposite of a circular 'prediction' of the final gains. The stability remark in Sec. 3.3 is explicitly conditional ('If the PAC-trained drift approximates the residual...') and is an unproven assumption rather than a self-referential derivation. References [6] and [28] are external works, not self-citations by the present authors, and no load-bearing argument invokes an author-generated uniqueness theorem. The uncertainty-calibration analysis in Sec. 4.4 is a post-hoc diagnostic using the model's own scores, not a fitted parameter renamed as a prediction. The fact that the bridge conditions on the target video distribution at inference is a potential correctness/attribution concern about late interaction versus bi-encoder comparisons, but it is not circularity: the cost matrix in Eq. (10) explicitly defines the pairwise cost in terms of the bridged text distribution and the target video distribution, so the paper does not silently equate a fitted quantity with a prediction. Overall, no circular step can be exhibited from the paper's equations or citations.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

No new physical entities, forces, or dimensions are postulated. The Distribution Bridge and probabilistic heads are architectural components built from standard network blocks, not invented entities in the sense of the ledger.

free parameters (6)
  • Temperature τ
    Used in the PAC contrastive loss (Eq. 12); its value is never reported, yet it controls the entire ranking sharpness.
  • χ_t normalization range = [0.2, 0.6]
    Step-size schedule normalized to this range (Sec. 3.3); the normalization constant is hand-picked and affects convergence.
  • β_s schedule endpoints = βmin=1e-4, βmax=2e-2
    Linear noise schedule following DITS; chosen by hand in Sec. 3.3.
  • Truncated refinement steps T' = 32 (of N=1000)
    Selected for accuracy/efficiency trade-off on MSR-VTT (Table 4); default value.
  • Log-variance clip and bias init = clip [−6, 2]; bias init −2.0
    Stability hyperparameters (Sec. 3.2); the init encourages low initial uncertainty.
  • Training budget = 10 epochs, batch 32, LR 1e-6/1e-4, dropout 0.3/0.4
    Main-table results use 10 epochs while ablations use 5; baseline numbers are not re-run under the same budget, so part of the gain may be training-time.
axioms (4)
  • domain assumption Diagonal-Gaussian encoding of embeddings captures semantic uncertainty
    The paper models text/video as N(µ, diag(σ²)) and claims variance reflects ambiguity; no independent evidence that this parametrization captures the claimed uncertainty beyond the self-reported diagnostics.
  • ad hoc to paper PAC-trained drift approximates x_t − x* so the refinement is a contraction
    Sec. 3.3 states this as a conditional stability intuition ('If the PAC-trained drift approximates the residual...'); no proof or empirical check that the learned drift satisfies this.
  • standard math Closed-form KL divergence for diagonal Gaussians
    Eq. 11 is a standard result.
  • domain assumption Directional KL as a ranking-sensitive cost is better than W2
    Supported only by the single ablation in Table 5(b); no theoretical grounding.

pith-pipeline@v1.3.0-alltime-deepseek · 11828 in / 13973 out tokens · 138635 ms · 2026-08-01T08:46:54.345728+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Distribution-Alignment Bridge for Uncertainty-Aware Text-to-Video Retrieval." pith.science (2026). https://pith.science/paper/ALO456CD

@misc{pith2026260720984,
  author       = {Pith},
  title        = {Pith review of: Distribution-Alignment Bridge for Uncertainty-Aware Text-to-Video Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ALO456CD}},
  note         = {Machine review of arXiv:2607.20984}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

This paper proposes the Distribution-Alignment Bridge (DAB), a framework that reconceptualizes text-to-video retrieval as a distribution alignment task rather than traditional deterministic point matching. By modeling both text and video embeddings as Gaussian distributions defined by mean and variance, DAB explicitly accounts for modality-specific uncertainty. We employ a deterministic, diffusion-inspired bridge to iteratively refine text distributions toward their target video distributions through a truncated refinement process. This approach unifies probabilistic embedding and distributional transformation into a cohesive, end-to-end trainable system. To optimize cross-modal similarity, we introduce a distribution-aware contrastive loss based on Kullback-Leibler divergence. Extensive evaluations on MSR-VTT, MSVD, and VATEX benchmarks confirm that DAB significantly outperforms existing probabilistic and diffusion-based baselines, while providing calibrated uncertainty-aware ranking through bridge-induced distributional margins.

Figures

Figures reproduced from arXiv: 2607.20984 by Jihoon Lee, Kyeongmo Chae, Sangtae Ahn.

Figure 1
Figure 1. Figure 1: Overall architecture of the proposed Distribution-Alignment Bridge [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Probabilistic Embedding Module. An input embedding is normalized and linearly projected, then split into two heads that predict the mean (\mu ) and variance (\sigma ^{2} ) of a Gaussian distribution. For video inputs, a token-based uncertainty branch refines the variance by applying self-attention and mean pooling over frame tokens. The fused variance and predicted mean form the final distribution \protect… view at source ↗
Figure 3
Figure 3. Figure 3: Detailed structure of the Distribution Bridge module. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Uncertainty calibration on MSR-VTT. Queries are ranked by the bridge￾induced KL margin between the best and competing retrieved candidates, and retrieval accuracy is evaluated under selective prediction. Larger margins correspond to more reliable retrieval decisions, yielding substantially lower risk and AURC than random ordering. This indicates that DAB provides a calibrated confidence signal through dist… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative local-neighborhood example. For the query “a little girl does gymnastics”, DAB ranks the annotated ground-truth video first and retrieves multiple gymnastics-related clips near the top. The highlighted top-ranked region provides an instance-level illustration of the local semantic-neighborhood behavior quantified in [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 7 linked inside Pith

  1. [1]

    In: Proceedings of the 49th annual meeting of the association for computational linguistics: human language technologies

    Chen, D., Dolan, W.B.: Collecting highly parallel data for paraphrase evaluation. In: Proceedings of the 49th annual meeting of the association for computational linguistics: human language technologies. pp. 190–200 (2011)

    2011

  2. [2]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Chen, S., Zhao, Y., Jin, Q., Wu, Q.: Fine-grained video-text retrieval with hierar- chical graph reasoning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10638–10647 (2020)

    2020

  3. [3]

    arXiv preprint arXiv:2305.18171 (2023)

    Chun, S.: Improved probabilistic image-text representations. arXiv preprint arXiv:2305.18171 (2023)

    Pith/arXiv arXiv 2023

  4. [4]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Chun, S., Oh, S.J., De Rezende, R.S., Kalantidis, Y., Larlus, D.: Probabilistic embeddings for cross-modal retrieval. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8415–8424 (2021)

    2021

  5. [5]

    Advances in neural information processing systems34, 8780–8794 (2021)

    Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems34, 8780–8794 (2021)

    2021

  6. [6]

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

    Fang, B., Wu, W., Liu, C., Zhou, Y., Song, Y., Wang, W., Shu, X., Ji, X., Wang, J.: Uatvr: Uncertainty-adaptive text-video retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13723–13733 (2023)

    2023

  7. [7]

    Gorti, S.K., Vouitsis, N., Ma, J., Golestan, K., Volkovs, M., Garg, A., Yu, G.: X- pool:Cross-modallanguage-videoattentionfortext-videoretrieval.In:Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5006–5015 (2022)

    2022

  8. [8]

    Hao, X., Zhang, W.: Uncertainty-aware alignment network for cross-domain video- textretrieval.AdvancesinNeuralInformationProcessingSystems36,38284–38296 (2023)

    2023

  9. [9]

    Advances in neural information processing systems33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

    2020

  10. [10]

    Advances in neural information processing systems35, 8633– 8646 (2022)

    Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. Advances in neural information processing systems35, 8633– 8646 (2022)

    2022

  11. [11]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Hur, C., Hong, J.h., Lee, D.h., Kang, D., Myeong, S., Park, S.h., Park, H.: Narrat- ing the video: Boosting text-video retrieval via comprehensive utilization of frame- level captions. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 24077–24086 (2025)

    2025

  12. [12]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Jeong, B., Park, J., Kim, S., Kwak, S.: Learning audio-guided video representation with gated attention for video-text retrieval. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 26202–26211 (2025)

    2025

  13. [13]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Jin, P., Li, H., Cheng, Z., Li, K., Ji, X., Liu, C., Yuan, L., Chen, J.: Diffusion- ret: Generative text-video retrieval with diffusion model. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2470–2481 (2023) Distribution-Alignment Bridge 17

    2023

  14. [14]

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

    Lan, B., Xie, R., Zhao, R., Sun, X., Kang, Z., Yang, G., Li, X.: Hybrid-tower: Fine-grained pseudo-query interaction and generation for text-to-video retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 24497–24506 (2025)

    2025

  15. [15]

    arXiv preprint arXiv:2012.04124 (2020)

    Lee, S., Yu, Y., Kim, G., Breuel, T., Kautz, J., Song, Y.: Parameter effi- cient multimodal transformers for video representation learning. arXiv preprint arXiv:2012.04124 (2020)

    Pith/arXiv arXiv 2012

  16. [16]

    arXiv preprint arXiv:1711.05101 (2017)

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)

    Pith/arXiv arXiv 2017

  17. [17]

    Neuro- computing508, 293–304 (2022)

    Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neuro- computing508, 293–304 (2022)

    2022

  18. [18]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Miech, A., Zhukov, D., Alayrac, J.B., Tapaswi, M., Laptev, I., Sivic, J.: Howto100m: Learning a text-video embedding by watching hundred million nar- rated video clips. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2630–2640 (2019)

    2019

  19. [19]

    In: Proceedings of the 2018 ACM on international conference on multimedia retrieval

    Mithun, N.C., Li, J., Metze, F., Roy-Chowdhury, A.K.: Learning joint embedding with multimodal cues for cross-modal video-text retrieval. In: Proceedings of the 2018 ACM on international conference on multimedia retrieval. pp. 19–27 (2018)

    2018

  20. [20]

    arXiv preprint arXiv:2010.02824 (2020)

    Patrick, M., Huang, P.Y., Asano, Y., Metze, F., Hauptmann, A., Henriques, J., Vedaldi, A.: Support-set bottlenecks for video-text representation learning. arXiv preprint arXiv:2010.02824 (2020)

    Pith/arXiv arXiv 2010

  21. [21]

    In: International conference on machine learning

    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. pp. 8748–8763. PmLR (2021)

    2021

  22. [22]

    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)

    Pith/arXiv arXiv 2022

  23. [23]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Reddy, A., Martin, A., Yang, E., Yates, A., Sanders, K., Murray, K., Kriz, R., De Melo, C.M., Van Durme, B., Chellappa, R.: Video-colbert: Contextualized late interaction for text-to-video retrieval. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 19691–19701 (2025)

    2025

  24. [24]

    Advances in neural information processing systems35, 36479–36494 (2022)

    Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text- to-image diffusion models with deep language understanding. Advances in neural information processing systems35, 36479–36494 (2022)

    2022

  25. [25]

    arXiv preprint arXiv:2209.14792 (2022)

    Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text- video data. arXiv preprint arXiv:2209.14792 (2022)

    Pith/arXiv arXiv 2022

  26. [26]

    arXiv preprint arXiv:2011.13456 (2020)

    Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score- based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)

    Pith/arXiv arXiv 2011

  27. [27]

    In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Wang, J., Sun, G., Wang, P., Liu, D., Dianat, S., Rabbani, M., Rao, R., Tao, Z.: Text is mass: Modeling as stochastic embedding for text-video retrieval. In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16551–16560 (2024)

    2024

  28. [28]

    Advances in Neural Information Processing Systems37, 3882–3906 (2024) 18 K

    Wang, J., Wang, P., Liu, D., Guan, Q., Dianat, S., Rabbani, M., Rao, R., Tao, Z.: Diffusion-inspired truncated sampler for text-video retrieval. Advances in Neural Information Processing Systems37, 3882–3906 (2024) 18 K. Chae et al

    2024

  29. [29]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Wang, X., Wu, J., Chen, J., Li, L., Wang, Y.F., Wang, W.Y.: Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4581–4591 (2019)

    2019

  30. [30]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Wang, Z., Sung, Y.L., Cheng, F., Bertasius, G., Bansal, M.: Unified coarse-to-fine alignment for video-text retrieval. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2816–2827 (2023)

    2023

  31. [31]

    Wu, W., Luo, H., Fang, B., Wang, J., Ouyang, W.: Cap4video: What can auxiliary captions do for text-video retrieval? In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10704–10713 (2023)

    2023

  32. [32]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5288–5296 (2016)

    2016