CosseratGen: Physics-Grounded Video Generation for Slender Deformable Objects

Controlling deformation in generated videos remains challenging for slender objects such as cables, hoses, and ribbons. Specifying an object's path does not determine how its shape should respond to material properties, loads, and constraints. The underlying challenge is to model the distributed dynamics that govern bending and rotation along the object. We introduce CosseratGen, a framework that combines learned slender body dynamics with pretrained video generation. At its core, the Direct-Time Neural Cosserat Operator (DT-NCO) learns from Cosserat rod simulations to predict centerlines and material frames given initial states and prescribed physical conditions, at requested times within the learned horizon. Material frames capture the orientation of local cross sections, distinguishing anisotropic bending responses that the initial centerline alone cannot determine. Projected predictions guide a frozen video generator, allowing physical conditions to control deformation while the scene prompt specifies appearance. Experiments demonstrate improved centerline prediction on withheld mechanics compositions and improved structural adherence in generated videos. By combining learned dynamics with generative priors, CosseratGen enables control over slender object deformation through explicit physical parameters.