Video generation models have become adept at producing visually compelling motion, but they systematically fail when it comes to fluid phenomena. According to a new arXiv preprint, outputs for smoke, fire, and explosion debris lack coherent vortical structure and collapse into bland dynamics.
The paper introduces FluidGen, a physics-aware approach that uses motion-guided post-training to address this shortcoming. Rather than training from scratch, the method post-trains existing models with motion guidance, steering them toward physically plausible fluid behavior.
The abstract does not provide implementation details or quantitative results, so the claims are limited to the described failure mode and the proposed remedy. The work highlights a specific weakness in current video generation and suggests a targeted fix.