PaperlayerView exact revision on arXiv

Source-linked AI summary

World Models That Know When They Don't Know - Controllable Video Generation with Calibrated Uncertainty

Zhiting Mei, Tenny Yin, Micah Baker, Ola Shorinwa, Anirudha Majumdar

arXiv:2512.05927v2cs.CVcs.AIcs.RO

TL;DR

Controllable video models can generate physically inconsistent future frames but lack confidence estimates, limiting fine-grained assessment of their outputs. C3 estimates calibrated uncertainty at subpatch resolution in latent space and maps it to pixel-space heatmaps; experiments report hallucination localization, calibrated estimates, and out-of-distribution detection.

  • Problem

    Controllable video models can hallucinate physically inconsistent future frames, yet lack confidence estimates, and existing video UQ provides only task-level estimates rather than dense frame-level uncertainty.

    Source: §1 · p. 2

  • Method

    C3 trains controllable video models for accuracy and calibration with proper scoring rules, estimates subpatch-level uncertainty in latent space, and maps it to interpretable pixel-space heatmaps.

    Source: §1 · p. 2

  • Results

    C3 localizes hallucinations, produces calibrated estimates across robot embodiments and tasks, and detects out-of-distribution inputs, including in real-world robot experiments.

    Source: §1 · p. 2, §5 · p. 11

  • Takeaways & Limitations

    C3’s confidence heatmaps provide interpretable pixel-space estimates that are reported to align with human intuition and highlight hallucinated regions.

    Source: §1 · p. 2, §5 · p. 11

    C3’s theoretical calibration guarantees apply only within the training distribution; observed calibration is influenced by training-data diversity and inference-time distribution shift.

    Source: §6 · p. 11

Abstract

from arXiv · show

Recent advances in generative video models have led to significant breakthroughs in high-fidelity video synthesis, specifically in controllable video generation where the generated video is conditioned on text and action inputs, e.g., in instruction-guided video editing and world modeling in robotics. Despite these exceptional capabilities, controllable video models often hallucinate - generating future video frames that are misaligned with physical reality - which raises serious concerns in many tasks such as robot policy evaluation and planning. However, state-of-the-art video models lack the ability to assess and express their confidence, impeding hallucination mitigation. To rigorously address this challenge, we propose C3, an uncertainty quantification (UQ) method for training continuous-scale calibrated controllable video models for dense confidence estimation at the subpatch level, precisely localizing the uncertainty in each generated video frame. Our UQ method introduces three core innovations to empower video models to estimate their uncertainty. First, our method develops a novel framework that trains video models for correctness and calibration via strictly proper scoring rules. Second, we estimate the video model's uncertainty in latent space, avoiding training instability and prohibitive training costs associated with pixel-space approaches. Third, we map the dense latent-space uncertainty to interpretable pixel-level uncertainty in the RGB space for intuitive visualization, providing high-resolution uncertainty heatmaps that identify untrustworthy regions. Through extensive experiments on large-scale robot learning datasets (Bridge and DROID) and real-world evaluations, we demonstrate that our method not only provides calibrated uncertainty estimates within the training distribution, but also enables effective out-of-distribution detection.

1 Introduction

C3 addresses the lack of fine-grained confidence estimates in controllable video models by predicting calibrated uncertainty at continuous, subpatch scales. Its latent-space approach also produces interpretable confidence heatmaps and detects out-of-distribution inputs in real-world robot experiments.

1 Introduction

  • C3 trains controllable video models for accuracy and calibration using proper scoring rules, with calibrated estimates demonstrated on Bridge and DROID.Calibration means the estimates are neither underconfident nor overconfident.Source: §1 · p. 2
  • Estimating uncertainty in latent space avoids pixel-space computation costs and enables subpatch-level detail finer than patch-level representations.The approach is applicable to latent-space video architectures without specialized adaptation.Source: §1 · p. 2
  • Confidence estimates are negatively correlated with generated-video error against ground truth, while heatmaps associate high uncertainty with hallucinations.The heatmaps decode latent uncertainty into interpretable pixel-space confidence estimates.Source: §1 · p. 2
  • C3 maps latent uncertainty to temporal RGB heatmaps that highlight untrustworthy video regions.The paper reports that higher-uncertainty areas are associated with hallucinations.Source: §1 · p. 2
  • C3 detects out-of-distribution environment conditions and actions on a WidowX 250 robot, even when distribution shift significantly compromises generated-video quality.The paper also reports calibrated estimates in these out-of-distribution settings.Source: §1 · p. 2

2 Related Work

Prior work advanced high-fidelity video generation and explored uncertainty quantification for image and video models, but video uncertainty estimates remained coarse. C3’s architecture uses a probe on video latents to estimate confidence alongside video generation.

2 Related Work

  • Recent high-fidelity video models use diffusion- or flow-based modeling, and later methods condition video generation on robot actions [1] [2] [3] [20] [23] [24] [25].Earlier approaches included local pixel transformations, GANs, and variational inference [12] [14] [16] [17] [19].Source: §2 · p. 2
  • C3’s UQ probe operates on video latents to estimate confidence during video generation; high-uncertainty regions are shown in red as hallucinations.The probe and video-generation components can also be trained independently.Source: Figure 2 · p. 3
  • Prior image-generation UQ used Bayesian variance decomposition, ensembles, or language descriptions, while video UQ remained difficult given spatio-temporality and computation costs [6].The existing video approach provides composite confidence per video rather than dense frame- or pixel-level estimates [6].Source: §2 · p. 3

3 Uncertainty Quantification of Video Models

C3 adds dense, subpatch-level confidence prediction to controllable video generation. It estimates confidence in latent space, trains predictions against accuracy using proper scoring rules, and decodes them into interpretable RGB heatmaps.

3.1 Confidence Prediction

  • C3 predicts confidence for each generated-video subpatch alongside action-conditioned video generation, framing uncertainty estimation as classification over video accuracy.A latent-space transformer probe estimates confidence during training and inference; the components can also be trained independently.Source: §3.1 · pp. 3–4

3.2 Model Architectures

  • The method defines accuracy from element-wise distance to ground truth and trains its confidence models with proper scoring rules.The paper states that predicted confidence is calibrated if the probe converges to an optimal solution.Source: §3.2 · pp. 4–5
  • Fixed-scale classification predicts accuracy at one chosen threshold, while multi-class classification discretizes prediction accuracy into confidence bins.Fixed-scale models are typically faster to train but less general across accuracy resolutions; multi-class predictions use softmax and cross-entropy.Source: §3.2 · pp. 4–5
  • Continuous-scale binary classification conditions confidence prediction on an inference-time accuracy threshold, with training sampling thresholds and adaptively discretizing them for efficiency.When domain knowledge is available, the paper says more informed discretization or sampling schemes can improve training efficiency.Source: §3.2 · p. 5

3.3 Decoding Latent Confidence Predictions

  • C3 maps latent confidence into RGB space by interpolating among latent encodings of red-, green-, and blue-only frames, then decoding with the video tokenizer.The resulting uncertainty heatmaps are reported to identify generated-video areas containing hallucinations.Source: §3.3 · pp. 5–6

4 Experiments

Experiments evaluate C3’s calibration, interpretability, out-of-distribution detection, and ablations on robot-manipulation video models. Results show calibrated estimates, uncertainty localized to hallucinations and OOD conditions, and no measured video-quality penalty from adding UQ.

4.1 Are C3’s uncertainty estimates calibrated?

  • All three C3 architectures produce well-calibrated uncertainty estimates; calibration varies modestly with the tradeoff between continuous-scale flexibility and fixed-scale calibration.Figure 3 reports average ECE and MCE across test videos; CS-BC averages errors across ten thresholds, FSC uses its training scale, and MCC averages across classes.Source: §4.1 · pp. 6–7, Figure 3 · p. 6

4.2 Are C3’s uncertainty estimates interpretable?

  • C3’s confidence heatmaps mark physically implausible pot-object hallucinations as highly uncertain, while confidence is negatively correlated with video error for FSC and CS-BC.For FSC and CS-BC, the correlations are −0.373 and −0.172, respectively, at 99% significance; MCC’s correlation is positive under its original binning because rightmost bins receive inadequate supervision.Source: §4.2 · pp. 7–8

4.3 Detecting OOD Inputs at Inference

  • In real-world OOD tests, C3 localizes uncertainty around unfamiliar background objects and captures uncertainty that increases over time under unseen lighting.The Bridge setup tests five OOD axes using 50 ground-truth trajectories; Figure 8 shows examples from the background and lighting categories.Source: §4.3 · pp. 8–9, Figure 8 · p. 10

4.4 Ablations

  • C3 marginally improves perceptual metrics and wins vanilla-model head-to-head rankings with scores of 62.5%, 58.8%, and 62.5% on SSIM, PSNR, and LPIPS.The reported comparison indicates that adding UQ does not degrade video quality.Source: §4.4 · pp. 9–10, Table 1 · p. 10
  • L1 latent-space error is strongly correlated with SSIM, PSNR, and LPIPS at the 99% significance level, without decoding videos to compute those metrics.This supports using latent error as an efficient proxy aligned with standard perceptual metrics.Source: §4.4 · p. 10, Table 2 · p. 10
  • Removing stop-gradient yields calibration similar to the stop-gradient variant, with ECE and MCE differences of about 5e−3 and 3e−3, but adds computation overhead.Stop-gradient provides a computational edge, especially for video models with billions of parameters.Source: §4.4 · p. 10

5 Conclusion

C3 trains controllable video models to estimate calibrated uncertainty, addressing hallucinations in generated video. Its latent-space estimates and pixel-space heatmaps localize untrustworthy regions, and experiments demonstrate calibration and OOD detection across robot embodiments and tasks.

  • C3 uses proper scoring rules to train video models for both accuracy and calibrated uncertainty.Source: §5 · p. 11
  • Estimating uncertainty in latent space avoids computational challenges and training instability associated with pixel-space approaches, while mapping it to pixel-space confidence estimates supports intuitive visualization.Source: §5 · p. 11
  • C3 localizes hallucinations and demonstrates calibrated confidence across robot embodiments and tasks, as well as OOD detection at inference.Source: §5 · p. 11

6 Limitations and Future Work

C3’s theoretical calibration guarantees apply only within the training distribution, while observed OOD calibration depends on training-data diversity and inference-time shift. Long-duration confidence consistency is also limited by the video model’s conditioning history.

6 Limitations and Future Work

  • C3’s theoretical calibration guarantees hold only within the training data distribution, and observed calibration depends on training diversity and inference-time distribution shift.The authors report calibrated estimates in OOD settings, but future work will explore training strategies for broader test-distribution coverage.Source: §6 · p. 11
  • Long-duration confidence consistency is limited by the conditioning history; shorter histories may cause C3 to lose track of uncertain patches over time.The authors identify long-duration video generation as an open research problem.Source: §6 · p. 11
  • The authors identify rigorous uncertainty quantification as important to the practical effectiveness of maturing robotics applications of video models.Source: §6 · p. 11

A Detecting OOD Inputs at Inference

In real-world OOD conditions, C3 highlights regions associated with hallucinations, including incorrect robot-object interactions, unfamiliar dynamics, and lighting-driven recoloring.

A Detecting OOD Inputs at Inference

  • C3 marks hallucinations under clutter, unfamiliar end-effector objects, and extreme lighting as high-uncertainty regions.Examples include inaccurate robot-object interactions, unfamiliar robot dynamics, and scene recoloring; Figure 10 depicts the extreme-lighting case.Source: A Detecting OOD Inputs at Inference · p. 15, Figure 10 · p. 15
  • Under OOD environments, C3 remains well-calibrated with only a very small calibration drop from nominal conditions, with ECE 9.98e−2 and MCE 1.71e−1.Source: A Detecting OOD Inputs at Inference · p. 15

B Calibration on the Bridge Dataset

On Bridge, C3 produces well-calibrated confidence estimates across models and accuracy thresholds. FSC calibrates best but uses a fixed threshold, while MCC and CSBC offer broader flexibility.

B Calibration on the Bridge Dataset

C Evaluations on the DROID Dataset

On DROID, C3 provides calibrated confidence estimates across diverse tasks and camera views, and its confidence aligns with video accuracy. It also localizes uncertain regions in degraded multi-view generations.

C Evaluations on the DROID Dataset

  • DROID test-set calibration errors are ECE 7.28e−2 and MCE 1.74e−1, and C3 is reported well-calibrated across diverse robotics prediction problems.The results include multi-view inputs and different robot embodiments; Figure 15 shows the reliability diagram.Source: C Evaluations on the DROID Dataset · p. 17, Figure 15 · p. 17
  • On DROID, confidence and absolute latent-video error have a negative correlation of −0.149, significant at greater than 99%.The authors interpret this as greater uncertainty when generated videos are more likely wrong.Source: C Evaluations on the DROID Dataset · p. 17
  • Despite lower video quality in multi-view DROID generations, C3 localizes uncertain regions, including a morphing gripper and a blurry wrist-camera background.Figure 16 shows side-camera and wrist-camera views.Source: C Evaluations on the DROID Dataset · p. 17, Figure 16 · p. 18

D Additional Ablations

The ablations compare scoring rules, diffusion forcing, and additional uncertainty baselines. C3 remains calibrated with BCE and Brier scores, while diffusion forcing degrades calibration; C3 is more competitive than the ensemble at lower computational cost and outperforms the heuristic baseline.

Proper Scoring Rules

  • The BCE and Brier variants differ by about 3e−4 in ECE and 6e−3 in MCE, with similar calibration across confidence bins.Both models are well-calibrated and more conservative when unsure; the reliability diagrams show similar trends.Source: D Additional Ablations · p. 19, Figure 17 · p. 19

Diffusion forcing

  • With diffusion forcing, ECE rises to about 3.3e−1 and MCE to about 5.54e−1, degrading C3’s calibration.The authors hypothesize recurrence increases confidence conservatism; a comprehensive analysis is left to future work.Source: Figure 18 · p. 19, D Additional Ablations · p. 19

Additional Baselines

  • C3 outperforms the heuristic uncertainty baseline and is more competitive than the ensemble at lower computational cost while achieving comparable calibration.The baselines use correlation with latent-space error because they do not estimate uncertainty with probability distributions, so classification calibration metrics do not apply.Source: Table 3 · p. 20, D Additional Ablations · p. 20

E Additional Implementation Details

The implementation trains and evaluates video models on Bridge and a DROID subset, and visualizes latent-space errors in RGB. The preliminaries describe diffusion-based generation and proper scoring rules used to assess calibration.

E Additional Implementation Details

  • Training uses the full Bridge train split and a DROID subset covering success and failure videos across tasks and environments; evaluation uses 110 Bridge and 83 DROID test trajectories.Models train for 50k iterations at 256 × 256 resolution; inference evaluates 10 εv values from 0.1 to 1.0, and calibration errors use 20 bins.Source: E Additional Implementation Details · p. 20
  • An error span of [0, 1] captures essentially all observable latent-space errors, with almost no green indicating the maximum-error end of the span.The RGB color map uses blue for minimum error, green for maximum error, and red for the middle region.Source: E Additional Implementation Details · p. 20, Figure 19 · p. 21

F.1 Video Diffusion/Flow-Based Models

  • Video diffusion models learn a video distribution by adding noise in a forward process and restoring structure in a reverse process; DDIMs accelerate generation by using shorter reverse timesteps.The model uses velocity prediction, and accuracy can be expressed in velocity space to avoid executing the reverse process during training.Source: §F.1 · pp. 21–22
  • The robot dataset is organized as trajectories pairing current observations and actions with next observations, and inference samples new video frames.The input conditions include video frames and action sequences.Source: §F.1 · p. 22

F.2 Proper Scoring Rule

  • A proper scoring rule penalizes predictions that differ from the true probability distribution, with equality uniquely minimized at the true distribution for a strictly proper rule.The preliminaries identify Brier score, cross entropy, and binary cross entropy as examples, and ECE/MCE as calibration-deviation metrics.Source: §F.2 · p. 22

G Proofs

The proof derives calibration of predicted confidence from minimizing a proper scoring rule. Its result depends on the confidence-probe parameters converging to an optimal solution.

Uncertainty Decomposition

  • When ϕ converges to an optimum, the predicted confidence ˆq is calibrated for the video diffusion model’s generated video.The proof uses the proper-scoring-rule property to establish the result.Source: G Proofs · p. 23
  • At convergence, P[Y = 1 | Q = ˆq] = E[Y | Q = ˆq] = ˆq, expressing calibration of predicted confidence.The proof identifies this equality as the calibration result.Source: G Proofs · p. 23
Loading 2512.05927v2…