CompAdapt addresses the gap between visual realism and physical correctness in text-to-video generation. It builds a language-to-dynamics-to-video pipeline where a text-to-physical parser turns prompts into motion types, temporal relations, and initial physical parameters. Neural ODE dynamics modules predict trajectories, and a composite motion module combines parallel or multi-stage motions while a multi-object interaction module handles collisions.
The framework also supports one-shot adaptation: dynamics-aware prior matching selects the most compatible dynamics module from a single observation, so it can handle novel environments like lunar gravity or low-friction surfaces without retraining the core model. A physics-aware latent feature fusion module preserves visual fidelity under fast and complex motion. According to the authors, CompAdapt improves physical consistency over general T2V and physics-constrained baselines on physics-focused benchmarks while maintaining visual quality and adaptability.