PhysVid: Physics Aware Local Conditioning for Generative Video Models

TL;DR

PhysVid introduces a physics-aware local conditioning method for generative video models, significantly enhancing physical plausibility by incorporating chunk-based physics descriptions and negative prompts during inference.

Contribution

The paper proposes a novel local conditioning scheme with physics-grounded annotations and negative prompts, improving physical consistency in generated videos.

Findings

PhysVid improves physical commonsense scores by approximately 33% on VideoPhy.

It increases physical plausibility by up to 8% on VideoPhy2.

The approach effectively steers video generation away from law violations.

Abstract

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 $\approx 33\%$ over baseline video generators, and by up to $\approx 8\%$ 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.