PaperlayerView exact revision on arXiv

Source-linked AI summary

How Confident are Video Models? Empowering Video Models to Express their Uncertainty

Zhiting Mei, Ola Shorinwa, Anirudha Majumdar

arXiv:2510.02571v1cs.CVcs.AIcs.CL

TL;DR

Video models can hallucinate without expressing uncertainty, and the paper addresses the absence of video-model UQ. It introduces a calibration metric, the latent-modeling method S-QUBED, and a benchmarking dataset; experiments report calibrated uncertainty estimates negatively correlated with task accuracy and effective estimation of both uncertainty components.

  • Problem

    Video models can hallucinate, yet the paper reports that no method existed to quantify their uncertainty, despite the safety concern.

    Source: §1 · p. 2

  • Method

    The paper proposes a calibration metric, black-box latent-modeling method S-QUBED for separating aleatoric and epistemic uncertainty, and a video-UQ dataset.

    Source: §1 · p. 2

  • Results

    Experiments report calibrated total uncertainty negatively correlated with task accuracy and effective estimation of aleatoric and epistemic uncertainty.

    Source: §1 · p. 2, §5.4 · p. 9, §6 · p. 9

  • Takeaways & Limitations

    The work provides a framework and dataset for evaluating and benchmarking uncertainty quantification in video models.

    Source: §1 · p. 2, §6 · p. 9

    S-QUBED requires generating multiple videos to estimate epistemic uncertainty, adding computational overhead.

    Source: §7 · p. 10

Abstract

from arXiv · show

Generative video models demonstrate impressive text-to-video capabilities, spurring widespread adoption in many real-world applications. However, like large language models (LLMs), video generation models tend to hallucinate, producing plausible videos even when they are factually wrong. Although uncertainty quantification (UQ) of LLMs has been extensively studied in prior work, no UQ method for video models exists, raising critical safety concerns. To our knowledge, this paper represents the first work towards quantifying the uncertainty of video models. We present a framework for uncertainty quantification of generative video models, consisting of: (i) a metric for evaluating the calibration of video models based on robust rank correlation estimation with no stringent modeling assumptions; (ii) a black-box UQ method for video models (termed S-QUBED), which leverages latent modeling to rigorously decompose predictive uncertainty into its aleatoric and epistemic components; and (iii) a UQ dataset to facilitate benchmarking calibration in video models. By conditioning the generation task in the latent space, we disentangle uncertainty arising due to vague task specifications from that arising from lack of knowledge. Through extensive experiments on benchmark video datasets, we demonstrate that S-QUBED computes calibrated total uncertainty estimates that are negatively correlated with the task accuracy and effectively computes the aleatoric and epistemic constituents.

1 Introduction

Text-to-video models can hallucinate without expressing uncertainty, despite the safety stakes of outputs that misalign with user intent or physical laws. The paper introduces a video-UQ framework comprising a calibration metric, the black-box S-QUBED method, and a dataset for benchmarking.

1 Introduction

  • Video models can generate outputs misaligned with user intent or physical laws without expressing uncertainty, posing a safety concern.The paper illustrates this with a prompt for Jeff Einstein that produces Albert Einstein instead.Source: §1 · p. 2
  • The framework introduces a calibration metric, black-box S-QUBED, and a video-UQ dataset to support evaluation and benchmarking.The metric assesses alignment between uncertainty estimates and task accuracy; S-QUBED separates aleatoric from epistemic uncertainty.Source: §1 · p. 2
  • S-QUBED uses latent modeling to distinguish uncertainty from ambiguous prompts from uncertainty due to the model’s lack of knowledge.The method is designed for black-box video-generation models, including closed-source models.Source: §1 · p. 2
  • The authors curate a dataset of about 40K videos across diverse tasks to facilitate benchmarking uncertainty-quantification methods for video models.The data are generated using Cosmos-Predict2 [27].Source: §1 · p. 2

2 Related Work

Prior uncertainty-quantification work spans training-free and training-based approaches, while generative image-model methods have faced computational challenges. The paper situates video-model uncertainty quantification as an unaddressed area.

2 Related Work

  • Deep-learning UQ includes training-free methods such as perturbation, dropout, and test-time augmentation, and training-based Bayesian methods such as variational inference, Monte-Carlo Dropout, and ensembles.Training-based methods can be computationally expensive, limiting their application to large generative models.Source: §2 · p. 2
  • Prior generative-image UQ includes weight-based diffusion approaches and PUNC, which maps generated images into language for similarity-based uncertainty estimation.The cited diffusion approaches carry training or inference overhead, while PUNC is not applicable to video models.Source: §2 · pp. 2–3
  • The authors describe their work as the first exploration of UQ for video world models.The paper distinguishes its video focus from prior image-model UQ approaches.Source: §2 · p. 3

3 Problem Formulation

The paper formulates text-conditioned video generation as an unknown stochastic model mapping a prompt to a video. It treats uncertainty in the generated output as arising from multiple sources and groups it into aleatoric and epistemic categories.

3 Problem Formulation

  • The setting is black-box text-conditioned video generation, where an unknown stochastic model maps an input prompt to a generated video.The prompt is denoted ℓ and the video v.Source: §3 · p. 3
  • Uncertainty in a generated video can arise from prompt vagueness, randomness in real-world physical evolution, or limited training data.The paper focuses on quantifying total predictive uncertainty associated with the output.Source: §3 · p. 3
  • The paper broadly classifies predictive uncertainty into aleatoric and epistemic components.The formulation concerns uncertainty associated with the sampled video output.Source: §3 · p. 3

4 Uncertainty Quantification of Generative Video Models

S-QUBED represents text-conditioned generation as a two-stage latent process and decomposes predictive entropy into aleatoric uncertainty over latent prompts and epistemic uncertainty over videos given a latent prompt. It estimates the two components from prompt refinements and generated samples.

4 Uncertainty Quantification of Generative Video Models

  • S-QUBED factors generation into sampling a latent prompt from p(Z | ℓ), then sampling a video from p(V | Z); the factorization assumes V and ℓ are conditionally independent given Z.The latent prompt can be interpreted as a more fully specified version of the input prompt, consistent with prompt refinement used by video models.Source: §4 · pp. 3–4
  • Total predictive uncertainty is decomposed into h(V | Z), the epistemic component, and h(Z | ℓ), the aleatoric component.The paper identifies this as a standard decomposition and points to a proof in Section A.2.Source: §4 · p. 4

4.1 Aleatoric Uncertainty

  • Aleatoric uncertainty models prompt vagueness through the entropy of p(Z | ℓ), which is independent of the generated video under the paper’s modeling assumption.More specific prompts have semantically closer compatible refinements and lower latent-distribution entropy than underspecified prompts.Source: §4.1 · p. 4
  • To estimate aleatoric uncertainty, S-QUBED uses an LLM to generate compatible, more-specific prompts, embeds them, and fits a VMF distribution to model p(Z | ℓ).PCA reduces embedding dimensionality to avoid numerical instability, and approximate closed-form methods estimate VMF parameters.Source: §4.1 · pp. 4–5

4.2 Epistemic Uncertainty

  • Epistemic uncertainty is the conditional entropy h(V | Z), representing model doubt from limited knowledge that can generally be reduced with additional training data.The paper illustrates this with a model trained on cats and dogs producing semantically inconsistent videos for a lion prompt.Source: §4.2 · pp. 5–6
  • S-QUBED estimates epistemic uncertainty by sampling latent prompts, generating videos for each, embedding the videos, fitting conditional VMF distributions, and averaging their entropies.The authors use Monte Carlo sampling for the expectation over latent prompts because direct computation is intractable in general.Source: §4.2 · p. 6

5 Experiments

The experiments assess calibration with rank correlations between uncertainty and task accuracy, then test S-QUBED’s total uncertainty and its aleatoric and epistemic components. Across two video datasets, higher S-QUBED uncertainty is associated with lower CLIP accuracy.

5.1 Evaluation Setup

  • Evaluation covers VidGen-1M and Panda-70M, sampling about 200 tasks per category across five broad categories.On Cosmos-Predict2’s text-to-image-to-video pipeline, S-QUBED samples 10 latent states and 10 videos per state.Source: §5.1 · pp. 6–7

5.2 How do we evaluate uncertainty calibration of video models?

  • At 99% significance, Panda-70M CLIP score is strongly negatively correlated with uncertainty, while other perceptual metrics are not statistically significant.The authors explain that CLIP captures semantic information that better reflects task accuracy than metrics more sensitive to visual changes.Source: Figure 3 · p. 7, §5.2 · pp. 7–8
  • Calibration-metric selection compares Kendall correlations for SSIM, PSNR, LPIPS, and CLIP using 10 Panda-70M tasks ranked by prompt vagueness; VidGen-1M is excluded because its prompts vary less in detail.Prompt vagueness serves as a proxy for aleatoric uncertainty.Source: §5.2 · p. 7

5.3 Are our uncertainty estimates calibrated?

  • 99% (Panda-70M) and 89.9% (VidGen-1M) confidence levels mark negative correlations between S-QUBED total uncertainty and CLIP accuracy.The figure also shows more accurate samples at low estimated uncertainty and less accurate samples at high estimated uncertainty.Source: Figure 4 · p. 8, §5.3 · p. 8

5.4 Can S-QUBED effectively estimate both aleatoric and epistemic uncertainty?

  • Component calibration is tested where the counterpart uncertainty is near zero, using samples with the lowest aleatoric or epistemic uncertainty.The visual examples also show lower accuracy at high estimated uncertainty, including difficulty predicting the specific “Behind the Gloves” logo compared with human-centric videos.Source: §5.4 · pp. 8–9
  • 94.5% and 98.3% (Panda-70M), and 92.3% and 91.7% (VidGen-1M), confidence levels mark negative accuracy correlations for aleatoric and epistemic uncertainty, respectively.The authors report that S-QUBED effectively decomposes total uncertainty into these two components.Source: Figure 5 · p. 9, §5.4 · p. 9

6 Conclusion

The paper concludes that its framework provides calibrated uncertainty estimates for video models and supports benchmarking UQ methods. Its experiments demonstrate total-uncertainty calibration and effective separation of aleatoric and epistemic uncertainty.

6 Conclusion

  • The experiments demonstrate calibrated uncertainty estimates and effective disentanglement of aleatoric and epistemic uncertainty.The framework includes a metric for assessing calibration and a method for estimating both uncertainty components.Source: §6 · p. 9
  • S-QUBED uses latent modeling to estimate aleatoric and epistemic uncertainty without limiting assumptions.The method is presented as a black-box UQ method for video models.Source: §6 · p. 9
  • An open-source video dataset is provided to benchmark uncertainty-quantification methods for video models.Source: §6 · p. 9

7 Limitations and Future Work

S-QUBED has computational overhead because estimating epistemic uncertainty requires multiple video generations. Future work will pursue more efficient sampling, broader benchmark data, and evaluation on additional open-source models.

7 Limitations and Future Work

  • Estimating epistemic uncertainty with S-QUBED requires generating multiple videos, creating computational overhead.Future work will explore more efficient sampling strategies, including sampling in the video model's latent space.Source: §7 · p. 10
  • The UQ dataset is currently limited to the two benchmark datasets considered in this work; future work will extend it to new datasets.Source: §7 · p. 10
  • Future work will examine S-QUBED on new open-source models as they become available.Source: §7 · p. 10

A.1 Evaluation Setup

The evaluation uses visual-fidelity and semantic-similarity metrics, with common video dimensions and durations for perceptual comparisons. Metric values are averaged across generated videos for each task.

A.1 Evaluation Setup

  • SSIM, PSNR, and LPIPS assess visual fidelity, while CLIP cosine similarity captures more semantic information.Source: §A.1 · p. 13
  • LPIPS is negated to convert it from an error metric into an accuracy metric.Source: §A.1 · p. 13
  • For perceptual metrics, videos are resized to the same spatial dimensions and longer videos are subsampled to match duration.Source: §A.1 · p. 13
  • Each task receives the mean metric value across its generated videos.Source: §A.1 · p. 13

A.2 Proofs

The proof decomposes total predictive uncertainty into epistemic uncertainty in video generation and aleatoric uncertainty from prompt vagueness. It derives the decomposition by introducing the latent state and splitting the entropy terms.

A.2 Proofs

  • Total predictive uncertainty, defined as h(V | ℓ), decomposes into h(V | Z) for epistemic uncertainty and h(Z | ℓ) for aleatoric uncertainty.The latter depends on prompt vagueness rather than the video model.Source: §A.2 · pp. 13–14
  • The proof expresses entropy using the joint latent-state and video distribution, then splits the logarithm into log(p(V | z)) and log(p(z | ℓ)).Source: §A.2 · p. 13
  • Applying the Fubini-Tonelli theorem yields terms recognized as entropies; h(V | Z) removes uncertainty due to prompt ambiguity.Source: §A.2 · pp. 13–14
Loading 2510.02571v1…