Despite achieving high visual fidelity, video models still suffer from physical hallucinations, limiting their reliability as world simulators. Existing methods leverage model-internal or external signals to improve physical plausibility, but their roles and limitations require a unified explanation. We propose a two-dimensional physical-hallucination risk framework that defines low-density sampling risk and high-density drifting risk. The first risk captures generations assigned low likelihood by the video model, for example, a ball suddenly teleporting despite the model favoring continuous motion, while the second risk captures physical inconsistencies in high-likelihood generations, for example, a ball bouncing progressively higher without additional energy. Our analysis explains how model-internal signals steer sampling toward higher-density regions of the learned distribution to address low-density sampling risk. However, higher model likelihood does not guarantee physical plausibility, motivating external guidance to address high-density drifting risk. Based on this framework, we develop Hallucination-Aware Sampler (HA-Sampler), which applies external physical guidance within latent refinement to generate physically plausible videos, and HA-Scorer, which derives a likelihood-based signal from the generator itself for evaluating generated videos using continuous normalizing flow (CNF), with coarse ODE integration and finite-difference divergence estimates. Experiments on Physics-IQ and VideoPhy benchmarks support our theoretical analysis and demonstrate that our methods outperform the compared sampling and scoring SOTA baselines.
Figures & tables
Figure 1: Overview of our two-dimensional physics hallucination risk framework. Left: Low-density sampling risk captures generations assigned low likelihood by the video model, while high-density drifting risk captures physical errors in high-likelihood generations. Middle: Model-internal physics-aware signals steer sampling toward higher-density regions of the learned distribution to reduce low-density sampling risk. Right: External physics-aware guidance helps correct the drifting between model-preferred generation and physical plausibility to address high-density drifting risk.
Figure 2: Overview of our methods. (a) HA-Sampler uses an external loss on the predicted clean video to guide the perturbation noise. The guided noise is combined with the original endpoint prediction before denoising continues. Grey dashed lines denote the flow of gradients. (b) HA-Scorer integrates the clean latent z0 to the noise endpoint z1 under the fixed base velocity field using a coarse Heun solver. Finite-difference Hutchinson estimator provides the divergence integral. In step 3, zt and zt+1 fall on the same velocity line.
Figure 3: Performance comparison across the two-dimensional risk. Each axis reports the fraction of 100 cases with lower corresponding risk than the Euler sampler.
Figure 4: Sampling results on 2D toy distributions. The stacked layers compare Euler sampler, P&P (risk-1 baseline), guidance based on the negative log-likelihood (NLL) under the true data distribution (risk-2 baseline), and HA-Sampler. The Euler sampler uses a large number of sampling steps to reduce discretization error. Purple reference curves indicate the training-data geometry.
Figure 5: Qualitative comparison of samplers on Physics-IQ. We compare sampling methods in a dynamic collision scene (left) and a static scene (right). HA-Sampler produces more physically plausible results than the base sampler and P&P in both cases.
Figure 6: Qualitative comparison of candidates selection on Physics-IQ. For each prompt, all scorer methods select from the same candidate pool generated by the base model. Each row shows three frames capturing key stages of motion in the video.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Sampler
BoN
Random
WMReward
GeoPhys
Qwen3-VL
HA-Scorer (ours)
Base sampler
2
27.74
28.84
26.57
27.73
28.99
4
27.45
30.75
26.50
28.64
30.77
8
26.38
30.51
25.67
28.18
30.88
P&P sampler
2
27.28
28.79
27.08
28.33
29.38
4
27.09
29.88
26.34
28.42
30.48
8
27.02
30.64
26.97
28.74
31.40
Appendix
Table 6: Best-of- N results with different candidate pool sizes. We report Physics-IQ Verified scores for selection methods using N∈{2,4,8} candidates from each sampler. Random denotes uniform random selection. Higher scores indicate better performance. The best and second-best scores in each row are bold and underlined, respectively.
Sampler
Risk Decomposition
Δ Risk/dim
Prior term
Volume term
Risk/dim ↓
Euler
1.10503
-2.54120
-1.43617
0.00000
C-TAG ( λ=1.25 )
1.22623
-2.76701
-1.54079
-0.10462
P&P
1.21343
-2.68283
-1.46940
-0.03323
CFG
1.21803
-2.70666
-1.48863
-0.05246
WMReward ( ∇ )
1.22506
-2.21974
-0.99468
+0.44149
Appendix
Table 7: Low-density sampling risk on 100 cases. The prior and volume terms sum to the low-density sampling risk. All quantities are reported per latent dimension. Δ Risk is measured relative to the paired Euler baseline without classifier-free guidance. We use H=8 Heun steps and P=2 shared Hutchinson probes. ∇ denotes gradient guidance.
Sampler
Risk ↓
Δ Risk
Base (CFG =5 )
0.414941
0.000000
P&P
0.415273
+0.000332
WMReward ( ∇ )
0.414746
-0.000195
HA-Sampler (ours)
0.414180
-0.000761
Appendix
Table 8: High-density drifting risk on 100 cases. We report mean WMReward surprise as a proxy for high-density drifting risk. Δ Risk is measured relative to the base sampler with CFG =5 . ∇ denotes gradient guidance.
Table 9: Sensitivity of HA-Scorer to integration steps and probe count. We evaluate the full 100-case set and a fixed stratified 25-case subset. We fix p=2 in (a), H=8 in (b), and δrel=0.0125 throughout. ρ denotes the Spearman rank correlation coefficient. NFE counts model forward evaluations per case.
Figure 7: Qualitative comparison of T2V generation on VideoPhy. For each prompt, we compare different samplers using three representative frames from each generated video. HA-Sampler improves physical plausibility in both cases and helps correct inconsistencies in the generated interactions.
Figure 8: Qualitative comparisons on Physics-IQ. Each example shows GT, Base, and Ours from top to bottom, with five frames arranged chronologically. Ours denotes videos generated with HA-Sampler and selected using HA-Scorer.
Figure 9: Additional qualitative comparisons on Physics-IQ. The layout follows Figure 8 .
Figure 10: Qualitative comparisons on VideoPhy. For each prompt, we show five frames from videos produced by the base model and HA-Sampler (Ours).
Modern video generators routinely fail at physical dynamics: objects float, trajectories violate gravity, contacts vanish. Standard denoising and flow-matching objectives fit visual data distributions but do not explicitly penalize such physical violations. Existing remedies can improve physical consistency, but typically add substantial inference or training cost. Candidate-selection methods generate and score multiple videos, while gradient-based world-model guidance repeatedly decodes and re-encodes intermediate estimates. Generator-internal refinement adds perturbation and re-denoising loops, whereas post-training requires curated data and additional optimization. We propose Off-Manifold Refinement (OMR), an inference-time method that instead injects world-model feedback directly into a single sampling trajectory. During scheduled middle ODE steps, we augment the generator velocity with the gradient of an adapter-space V-JEPA 2.1 surprise energy. This external correction can move the latent away from the uncorrected sampling trajectory and toward regions ranked as more physically plausible by the frozen predictor, after which the generator continues rendering from the corrected state. A small trained latent-to-embedding adapter keeps the gradient tractable at inference, and both the video generator and the world model remain frozen. On our fixed 400-prompt VideoPhy-2 detailed subset, OMR lifts the joint Semantic-Adherence-and-Physical-Commonsense metric from 47.0% to 52.0% (+5.0pp absolute, +10.6% relative) over the base Wan2.2-T2V-A14B sampler. On a separate fixed 50-prompt efficiency subset, it requires 1.71× the base runtime rather than the multiplicative cost of reward/search alternatives. Project page: https://itruonghai.github.io/omr.
Hai Nguyen-Truong, Tuan-Anh Vu, Dang Huynh
Fulbright University Vietnam · University of California, Los Angeles
Modern video generative models produce visually impressive results, yet frequently violate basic physical principles. We propose Proprio, a training-free framework that enables a frozen video generator to assess and improve the physical plausibility of its own outputs. Inspired by proprioception, the biological sense of one's own movement, Proprio treats the model's flow residual under controlled latent perturbations as a self-scoring signal. Samples that are better explained by the generator's learned dynamics induce smaller and more stable residuals. We aggregate this signal across timesteps and perturbations, focus it on motion-relevant regions with a dynamic spatiotemporal mask, and use it for best-of-N search, gradient-based self-refinement, or both. Across text-to-video and image-to-video benchmarks, Proprio consistently improves physical plausibility, outperforming VLM-based scoring, and external world-model baselines in several settings. With TurboWan2.2, Proprio improves Physics-IQ from 32.2 to 37.5 (+16.5%) and VideoPhy2-hard physical commonsense from 45.6 to 55.0 (+20.6%). Human evaluation further shows that raters prefer Proprio-selected or refined videos for physical plausibility in roughly two-thirds of comparisons. These results suggest that frozen video generators contain actionable internal signals for evaluating and improving the physical plausibility of their own outputs.
Mariam Hassan, Kaouther Messaoud, Wuyang Li +1
1 École Polytechnique Fédérale de Lausanne (EPFL) · Télécom Paris, IP Paris
Generative video models achieve high visual fidelity but often violate basic physical principles, limiting reliability in real-world settings. Prior attempts to inject physics rely on conditioning: frame-level signals are domain-specific and short-horizon, while global text prompts are coarse and noisy, missing fine-grained dynamics. We present PhysVid, a physics-aware local conditioning scheme that operates over temporally contiguous chunks of frames. Each chunk is annotated with physics-grounded descriptions of states, interactions, and constraints, which are fused with the global prompt via chunk-aware cross-attention during training. At inference, we introduce negative physics prompts (descriptions of locally relevant law violations) to steer generation away from implausible trajectories. On VideoPhy, PhysVid improves physical commonsense scores by ≈33% over baseline video generators, and by up to ≈8% on VideoPhy2. These results show that local, physics-aware guidance substantially increases physical plausibility in generative video and marks a step toward physics-grounded video models.