FracGen: Learning How Objects Stretch and Tear with Physics-Informed Video Generation

What it claims

FracGen generates stretch-to-tear fracture video from one image of an intact object. It is controlled by an applied tensile load, Young's modulus and Poisson's ratio, and two stretch thresholds that set when tearing starts and when it completes. Under the same force, changing only the thresholds produces early tearing or delayed tearing after larger deformation. No simulation runs at inference time. The authors call it the first model to generate the full deformation-to-fracture progression.

How it works

FracSim (training data). FracSim extends PhysGaussian, which couples 3D Gaussian splatting with the material point method (MPM), with a per-particle damage scalar d ∈ [0,1]:

• Damage rule. Damage is driven by the maximum principal stretch λ_max of the deformation gradient. It rises linearly from λ_onset to λ_crit.
• Irreversibility. Damage is a running maximum, so it never decreases during unloading.
• Feedback into the simulation. Damage scales stress as τ̃ = (1−d)τ before the momentum update.
• No plasticity. Damage is the only softening mechanism; plastic flow is not modelled.

Four per-particle fields are rendered through the Gaussian compositing path in place of colour: velocity, the Frobenius norm of Green–Lagrange strain, von Mises stress (from degraded Cauchy stress) and damage. This gives pixel-aligned maps with no extra simulation runs.

FracGen (the generator).

• Base model. Wan 2.1-1.3B-Control, adapted with LoRA. The 3D VAE stays frozen.
• Joint output. The RGB latent and the four map latents are concatenated along height, giving a canvas 5 frames tall. Following Chen et al. 2025, this adds no parameters.
• Conditioning. Signals enter channel-wise in Wan's 32-channel control layout:
• 16 channels: VAE-encoded moving Gaussian blobs that mark grip positions, mimicking a universal testing machine.
• 2 channels: E normalised as (log E − 6)/3 over 10³–10⁹ Pa, and ν over 0.1–0.5.
• 4 channels: a ±1 code for low- versus high-stretch fracture, plus λ_onset and λ_crit normalised from [1, 10].
• 10 channels: zero padding.
• Losses. Flow matching, plus a latent consistency loss that couples predicted strain, stress and damage. It is motivated by linear elasticity, which the paper calls an approximation.

flowchart LR
  A[Object image] --> D[Wan DiT + LoRA]
  B[Load blobs via VAE] --> D
  C[E, ν, λ_onset, λ_crit maps] --> D
  D --> E[RGB + flow/strain/stress/damage latents]
  E --> F[Flow-matching loss]
  E --> G[Strain-stress-damage consistency loss]

Evidence

The fetched text cuts off before any quantitative results. What is visible:

• A qualitative comparison on unseen objects (cake tearing, an extreme stretch of a Pikachu figure) against Force Prompting, PhyCo and CogVideoX.
• A claim that FracGen beats these on both physical and visual fidelity, measured on a new fracture benchmark with new metrics.

No benchmark numbers, dataset size or compute figures appear in the fetched text.

How to use it

The project site is https://fracgen.github.io. The fetched text gives no code, weights or licence. To reproduce, you would need:

• Wan 2.1-1.3B-Control.
• A PhysGaussian/MPM pipeline with the damage rule above.
• Pretrained 3DGS objects to simulate on.

Limits

• Loading. Only tensile, stretch-to-tear fracture driven by two opposing grips. Collision, bending and shattering are not addressed.
• Material model. No plasticity. Damage is a kinematic stretch proxy rather than an energy-based fracture criterion.
• Consistency loss. It assumes linear-elastic coupling in latent space, while the simulator uses nonlinear strain energy.
• Missing results. Quantitative results, ablations and the benchmark definition fall outside the fetched text. Check the full paper before relying on the performance claims.