JHU's FracGen teaches video models physics-accurate stretching and tearing from a single image

anand_bhattad · x · 2026-10-12

JHU researchers present FracGen, a physics-informed, controllable fracture-video generation framework: from a single image plus physical conditions (loading, material, fracture parameters), it generates plausible stretching and tearing videos.

Motivation: robots should anticipate how and when things break; understanding material failure also aids tougher material design and structural failure prediction.

Method:

Current video models can't even tear bread (it stretches like dough or morphs), while 3D-reconstruction-plus-simulation doesn't scale;
They built FracSim: MPM simulation via PhysGaussian augmented with a continuum damage model, rendering flow/strain/stress/damage maps paired with RGB video at no extra cost;
Wan is adapted with LoRA to jointly predict video and physical maps, plus a physics-consistency loss linking strain, stress and damage in latent space;
The model generalizes to unseen objects and materials (bread/cake tear along loading direction, rubber toys stretch far while staying connected), with fine control over tear location, crack speed and pre-failure deformation.

They also introduce a benchmark for physical plausibility of fracture videos, showing FracGen beats existing baselines on both physical and visual fidelity. The authors envision a data-model loop accelerating materials discovery and robot-world interaction.

Original post →

JHU's FracGen teaches video models physics-accurate stretching and tearing from a single image

More from Multimodal