Title: Enhancing Video Generation with Physics Constraints Through Mixture-of-Experts Latent Alignment

Large-scale video generation models have achieved significant strides in semantic coherence and visual fidelity, yielding clips that are increasingly realistic and consistent. However, these models often struggle with physical plausibility, as their pixel-level fitting approaches do not inherently respect the laws governing real-world motion and interaction. To bridge this gap, we introduce PILA (Physics-Informed Latent Alignment), a novel framework designed to integrate physics-structured latent guidance into the frozen flow-matching dynamics of pre-trained video models.

PILA operates by first utilizing anchored field estimation to transform the latents from a frozen generator into a structured physical attribute bank. This bank is organized into field-proxy slots, leveraging observable motion as a kinematic anchor to construct proxies for less directly observable physical properties. To manage the diverse nature of real-world dynamics, PILA implements a mixture-of-experts architecture across physical categories. It employs label-prior masked expert routing to select category-specific operator experts, with their refinements being regularized by operational residuals derived from physical relationships.

The final step involves fusing these refined proxies back into the physical attribute bank and decoding them into a correction for the flow-matching vector field. This process injects physics-aware guidance into the generation pipeline while maintaining the visual priors established by the pre-trained backbone. By employing staged adapter training on the Wan 2.1-1.3B model and directly transferring the learned adapter to Wan 2.2-14B, PILA attains state-of-the-art performance on VBench-2.0, VideoPhy-2, and PhyGenBench, excelling in both visual quality and benchmark-assessed physical plausibility.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC