LikePhys: Evaluating Intuitive Physics Understanding in Video Diffusion Models via Likelihood Preference

TL;DR

LikePhys presents a training-free, likelihood-preference framework for evaluating intuitive physics in video diffusion models by comparing model likelihoods on matched valid and invalid video pairs rendered with consistent appearance. It constructs a Blender-based benchmark of 12 scenarios across four physics domains and introduces PPE as a robust proxy aligned with human judgments, enabling zero-shot ranking across 12 pre-trained VDMs. The results show gradual improvements with modern architectures (notably DiT-based designs) but reveal persistent gaps in complex dynamics like fluid flows, underscoring the need for physics-aware objectives and longer temporal context. The method disentangles physics understanding from visual quality and offers a scalable, principled diagnostic to guide development of physically plausible video synthesis systems.

Abstract

Intuitive physics understanding in video diffusion models plays an essential role in building general-purpose physically plausible world simulators, yet accurately evaluating such capacity remains a challenging task due to the difficulty in disentangling physics correctness from visual appearance in generation. To the end, we introduce LikePhys, a training-free method that evaluates intuitive physics in video diffusion models by distinguishing physically valid and impossible videos using the denoising objective as an ELBO-based likelihood surrogate on a curated dataset of valid-invalid pairs. By testing on our constructed benchmark of twelve scenarios spanning over four physics domains, we show that our evaluation metric, Plausibility Preference Error (PPE), demonstrates strong alignment with human preference, outperforming state-of-the-art evaluator baselines. We then systematically benchmark intuitive physics understanding in current video diffusion models. Our study further analyses how model design and inference settings affect intuitive physics understanding and highlights domain-specific capacity variations across physical laws. Empirical results show that, despite current models struggling with complex and chaotic dynamics, there is a clear trend of improvement in physics understanding as model capacity and inference settings scale.

Figures (9)

• Figure 1: Overview of LikePhys. Our intuition is that a video diffusion model with a well learned underlying physics distribution should assign higher likelihoods to valid samples that obey physics laws and lower likelihoods to those invalid samples that violate them. We use Blender to create valid and invalid sample pairs through controlled physics parameters over multiple physics domains and scenarios (left). We then compare the diffusion model likelihood estimates over the constructed dataset to extract a quantitative intuitive physics understanding measure, the Plausibility Preference Error (PPE) (middle). Hence, we can compute a ranking of average PPE across pre-trained video diffusion models, that correlates with human preference. Lower values indicate stronger intuitive physics understanding (right).
• Figure 2: Method Overview. We prepare groups of videos via physics simulations with valid samples obeying physical laws and invalid samples containing deliberate violations. We then inject Gaussian noise into these videos and use a diffusion model to predict the noise and compute the denoising loss. For each valid–invalid pair, we compute a likelihood preference ratio that quantifies how the model favors physically plausible sequences, serving as a proxy for physics understanding.
• Figure 3: Evaluation benchmark. We organise 12 scenarios derived from four physical domains (Rigid Body Mechanics, Fluid Mechanics, Continuum Mechanics, Optical Effects) with their relative proportions (left). Rows from top to bottom (middle and right) show examples from four optical phenomena, rigid-body mechanics, continuum mechanics, and fluid mechanics. We display valid simulation (middle) and corresponding invalid variants (right) for physics violation in each domain.
• Figure 4: Analysis on influencing factors. Effect of model size, showing a steady overall decline in PPE (top left). Effect of training data size, where larger corpora generally yield lower error, though the correlation is weaker than that of model size (bottom left). Effect of context window size, showing consistent improvement in physics understanding as the window increases (top right). Effect of classifier-free guidance (CFG) strength, indicating that physics understanding remains largely stable across different strengths (bottom right).
• Figure 5: Analysis across physics domains and laws. On the left, we report detailed PPE across the four physics domains. Fluid Mechanics cases exhibit both the highest average error, while Rigid Body and Continuum Mechanics scenarios show moderate errors; Optical Effects cases lie in between. On the right, we map our domains to seven physics laws for a fine-grained analysis. Temporal continuity and conservation of energy show wide variation and higher median errors, whereas geometric invariance and optical consistency are handled more reliably.