Seeking Physics in Diffusion Noise

TL;DR

This paper investigates whether diffusion models encode physical plausibility signals and introduces a method to improve physical consistency during inference by selecting trajectories based on physics verification, reducing computational cost.

Contribution

It reveals that diffusion model features contain recoverable physics cues and proposes a trajectory selection method to enhance physical plausibility efficiently.

Findings

Physically plausible and implausible videos are partially separable in model features.

The proposed method improves physical consistency in generated videos.

Achieves comparable quality with fewer denoising steps, reducing inference cost.

Abstract

Do video diffusion models encode signals predictive of physical plausibility? We probe intermediate denoising representations of a pretrained Diffusion Transformer (DiT) and find that physically plausible and implausible videos are partially separable in mid-layer feature space across noise levels. This separability cannot be fully attributed to visual quality or generator identity, suggesting recoverable physics-related cues in frozen DiT features. Leveraging this observation, we introduce progressive trajectory selection, an inference-time strategy that scores parallel denoising trajectories at a few intermediate checkpoints using a lightweight physics verifier trained on frozen features, and prunes low-scoring candidates early. Extensive experiments on PhyGenBench demonstrate that our method improves physical consistency while reducing inference cost, achieving comparable results to Best-of-K sampling with substantially fewer denoising steps.