Seeking Physics in Diffusion Noise

2026-03-15

AI Summary

Existing video diffusion models often generate outputs lacking physical plausibility and struggle to balance efficiency with temporal consistency. This work identifies that the frozen intermediate denoising features of pretrained Diffusion Transformers (DiTs) contain discriminative signals indicative of physical reasonableness. Leveraging this insight, the authors propose a fine-tuning-free, inference-time trajectory selection strategy that integrates a lightweight physics validator, parallel denoising trajectory scoring, and an early pruning mechanism to progressively select physically consistent trajectories. Evaluated on the PhyGenBench benchmark, the method significantly improves the physical consistency of generated videos while reducing computational overhead, achieving performance comparable to Best-of-K sampling with substantially fewer denoising steps.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Search and Optimization: Sampling/Simulation-based Search
• Constraint Satisfaction and Optimization: Satisfiability Modulo Theories

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Security and Privacy: Data transparency and provenance

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.

Problem

• physical plausibility
• video diffusion models
• denoising trajectories
• physics verification

Innovation

• diffusion models
• physical plausibility
• trajectory selection
• feature separability
• inference efficiency