Video models separate risks to reduce physical hallucinations
Risk Decomposition for Hallucination-Aware Video Sampling
Computer Vision and Pattern Recognition
Summary
Video models can produce visually good videos but sometimes create impossible events, like objects teleporting or defying physics. The authors break down these issues into two kinds of risks: making unlikely scenes and drifting into physically inconsistent scenarios while still making likely videos. They show internal model cues help avoid unlikely scenes but do not prevent impossible physics, so they add external physical rules to guide video generation. Their new tools, HA-Sampler and HA-Scorer, help produce and evaluate more realistic videos, improving performance on standard physics video tests.
What this means in practice
- •For graphics developers: Generate videos with fewer physically impossible behaviors by integrating physical constraints during sampling.
- •For machine learning engineers: Evaluate generated video quality using likelihood-based physical consistency scores derived from the model itself.
Authors
Chenyan Zheng, Fan Wu, Ruibo Li, Jiaxing Li, Minghao Yang, Bang Zhang, Guosheng Lin
Abstract
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.