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

• The paper asks whether video generators can learn how objects break from limited data and lightweight fine-tuning, given that convincing fracture must capture deformation and failure progression.Source: §1 · p. 2
• FracGen generates tensile stretch-to-tear dynamics, where substantial deformation may precede failure and fracture behavior varies by material.Source: §1 · p. 2
• FracSim adds continuum damage mechanics to MPM and produces paired fracture videos and pixel-aligned physical maps from simulation state.The maps provide dense physical supervision without additional simulation runs beyond rendering.Source: §1 · p. 2
• FracGen jointly predicts RGB video and physical fields, with physics-informed losses encouraging consistency among predicted channels; the paper also introduces a fracture-video benchmark.Source: §1 · p. 2
• FracGen generalizes across unseen objects and outperforms existing baselines in both visual and physical fidelity.Source: §1 · p. 2

• Physics-informed 3D simulation couples continuum mechanics with scene representations or optimizes simulation parameters, while many solvers require manual scene specification and parameter tuning.Examples include PhysGaussian (Xie et al., 2024), Gaussian Splashing (Feng et al., 2025), PhysDreamer (Zhang et al., 2024), and DreamPhysics (Huang et al., 2025).Source: §2 · p. 3
• Physics-informed video generation uses explicit control signals or high-level reasoning and reward alignment, but the cited methods focus on rigid-body motion or simple deformation rather than material failure.The passage discusses Force Prompting (Gillman et al., 2025), PhysCtrl (Wang et al., 2025), PhyCo, DiffPhy (Zhang et al., 2025), and PhyGDPO (Cai et al., 2025).Source: §2 · p. 3
• The paper identifies FracGen as the first model to generate complex fracture dynamics across the full deformation-to-fracture progression.Source: §2 · p. 3

• The deformation gradient encodes local stretching, compression, rotation, and shearing, and polar decomposition separates rotation from stretch.Source: §3 · p. 3
• Strain measures local deformation and stress describes internal forces; a constitutive model relates them, while FracSim computes stress from strain energy with damage-dependent degradation.Source: §3 · p. 3
• MPM transfers particle quantities to a background grid to update particle states, while associated Gaussians deform with the material to render its evolving appearance.Source: §3 · p. 3
• The latent video generator maps RGB video to latents and uses flow matching to transport Gaussian noise toward video latents, conditioned on inputs.Source: §3 · p. 4
• The formulation is extended to generate fracture video and physical maps under specified material and force conditions; Figure 2 shows paired RGB and FracPhys outputs and control of fracture timing.Source: §3 · p. 4, Figure 2 · p. 4

4.1 FracSim: Simulating Fracture Video and Rendering FracPhys Maps

• FracSim integrates continuum damage mechanics into MPM to generate dynamic physics maps alongside RGB frames for training.The rendered maps include stress, strain, damage, and velocity, aligned spatially and temporally with each RGB frame.Source: §4.1 · pp. 4–5
• A scalar damage variable tracks accumulated failure from intact material to complete failure, and its degradation of stress feeds back into the MPM momentum update.Damage begins at an onset stretch, grows to complete failure at a critical stretch, and the running maximum preserves it during unloading.Source: §4.1 · p. 5

4.2 FracGen: Teaching Video Generator Fracture Dynamics

• FracGen jointly predicts RGB fracture video and FracPhys Maps, using flow-matching supervision and a latent consistency regularizer to couple strain, stress, and damage.The joint prediction supports inspection of the generated physical state without an external physics simulator.Source: §4.2 · pp. 5, 8, Figure 3 · p. 6
• FracGen conditions generation on applied forces, material properties, and fracture properties, while the intact object image supplies reference tokens.Applied loading is represented by opposing velocity signals; fracture properties encode low- versus high-stretch fracture categories.Source: §4.2 · p. 6, Figure 3 · p. 6
• The latent regularizer uses a damage gate to encourage stress latents toward a constitutive approximation, with greater regularization weight at lower noise levels.The gate provides a soft stress-degradation prior and does not assume VAE coordinates directly represent physical stress or damage.Source: §4.2 · p. 8

• The corpus samples 10 objects across low- and high-stretch materials; 6 objects yield 4,212 training samples, while 4 held-out objects provide 2,808 test samples.The test objects are evaluated against ground truth under varying force, material, and fracture conditions.Source: §5 · p. 8
• FracGen outperforms fine-tuned baselines across reported metrics, excelling in video quality and stretching-axis alignment; baselines miss realistic stretching and fracture progression.Table 1 reports quantitative results on the held-out test set, with all baselines fine-tuned on the same dataset.Source: Table 1 · p. 9, §5 · p. 9
• Evaluation pairs FVD, LPIPS, and PSNR with fracture-specific SAC, DPA, and constitutive consistency; DPA and consistency are used for ablations rather than baseline comparisons.SAC measures alignment of dominant deformation with the ground-truth stretching axis; constitutive consistency checks agreement between stress and strain in intact regions.Source: §5 · p. 9
• Generated videos and maps align under varying force directions, while low/high-stretch settings and stiffness changes produce different tearing behavior.The low- and high-stretch settings differ in stiffness and critical damage thresholds.Source: Figure 5 · p. 9, §5 · p. 10, Figure 6 · p. 10
• Adding joint FracPhys prediction (Stage 2) improves video quality, reducing FVD from 298 to 290, and yields pixel-aligned physical fields that enable direct inspection of generated dynamics (Tab. 2).Source: §5 · p. 10
• Constitutive loss changes FVD 290→266, DPA 0.438→0.536, and overall map MSE 0.0253→0.0175, while strain error rises slightly.The authors attribute the strain-error increase to prioritizing stress–strain coupling over per-map fidelity.Source: §5 · p. 10

6 Discussion

• FracGen adapts pretrained video generators to learn material-dependent stretching and tearing from simulation-derived supervision, jointly predicting physical fields whose evolution can be inspected.The authors suggest experimental validation could support material-response prediction and materials discovery.Source: §6 · p. 11
• The work covers tensile deformation and tearing, not brittle impact fracture or fragmentation; it inherits simulator errors, and multi-step diffusion sampling limits inference speed.Extending to other fracture types would require appropriate simulation data and failure models.Source: §6 · p. 11

AI use statement

• Generative AI assisted with language, literature, baseline setup, data reformatting, and the website, but not dataset generation, hypothesis or experiment design, or method implementation; authors reviewed outputs.Source: AI use statement · p. 11

• Stress, strain, and damage are normalized per video, clipped to [0, 1], and replicated across RGB channels; flow uses a Middlebury-style color wheel, with zero and background rendered black.Source: §A.2 · p. 14
• Each preprocessed physics-map video is encoded by the frozen 3D VAE, so the additional modality requires no architectural change.Source: §A.2 · p. 14

• MPM represents a body with particles carrying position, velocity, mass, and a deformation gradient that tracks accumulated deformation, including rotation and strain.Source: §A.3 · p. 14
• At each step, particle mass and momentum transfer to grid nodes, where velocities update under internal stress and external constraints.The particle-to-grid transfer uses an affine particle-in-cell scheme, with B-spline weights coupling particles to grid nodes.Source: §A.3 · pp. 14–15
• Updated grid velocities interpolate back to particles for advection, and each particle's deformation gradient evolves accordingly.Source: §A.3 · pp. 14–15

A.4 Additional Simulation and Training Details

• The stress-processing description reduces deviatoric Cauchy stress to a scalar von Mises equivalent stress.Figure 9 illustrates using point tracks to extract the principal stretching axis.Source: §A.4 · p. 15, Figure 9 · p. 17
• FracSim couples continuum damage to MPM and produces paired RGB fracture videos and dynamic physics maps under reproducible simulation configurations.Runs use 100 frames with 100 MPM substeps per frame and no gravity.Source: §A.4 · p. 15
• Two particle slabs on opposite sides of the object are driven at prescribed velocities along a selected tearing axis to impose tensile loading.The tearing axis spans 12 image-plane directions; speed and grip offset are varied.Source: §A.4 · p. 15
• Fixed-corotated elasticity is used for low-stretch objects, while high-stretch objects use compressible neo-Hookean elasticity.The latter supports stretches of 3× or more before failure.Source: §A.4 · p. 15
• Young’s modulus and Poisson’s ratio vary by class, shaping crack localization, necking, and cross-section thinning; rubber uses weaker velocity damping than bread-like objects.The per-class simulation parameters are summarized in Table 3.Source: §A.4 · p. 16, Table 3 · p. 16
• Damage accumulates from the largest principal stretch, and independently varied onset and critical stretch thresholds control weakening and fracture timing.The thresholds decouple when an object breaks from how it deforms beforehand.Source: §A.4 · p. 16
• Of 10 simulated objects, six are used for training and four for testing and benchmarking; 10 additional random object images are used for generalization tests.Training samples use three low-stretch and three high-stretch objects; the held-out set contains two of each.Source: §A.4 · pp. 16–17, Figure 8 · p. 17

A.5 Additional Details on Evaluation Metric Design

• The stretching-axis procedure extracts point tracks from a fracture video and estimates the dominant deformation axis over an early window, comparing it with the same quantity from a reference clip.It subtracts mean displacement from each tracked point’s displacement and drops points whose deformation-relative motion magnitude is below τ; if fewer than k survive, the axis is undefined.Source: §A.5 · pp. 17–18
• Stretching Axis Coherence compares prediction and reference principal stretching axes using angular deviation, with undefined predicted axes scored as failures.A Gaussian kernel maps angular error to a bounded similarity score; a perfect match scores 1 and the decay uses σ = 15°.Source: §A.5 · p. 18
• Damage Progression Alignment compares when prediction and reference reach checkpoints defined as fractions of each clip’s peak damage.Agreement is averaged over checkpoints from 0.1 to 0.9.Source: §A.5 · pp. 18–19
• Constitutive consistency measures how well predicted stress and strain in undamaged material follow a single linear stress–strain law through the origin.The metric fits one slope and records its coefficient of determination.Source: §A.5 · p. 19
• 0.902 versus 0.378: FracGen with constitutive loss achieves constitutive-consistency R²_cons of 0.902, while the variant without it reaches 0.378.Ground-truth maps score 0.948, treated as the upper bound.Source: §A.5 · p. 19

A.6 Additional Details on Building Annotation Tool

• The annotation tool lets users specify uniaxial load, material properties, and fracture properties for an input image.Figure 10 shows the interface.Source: §A.6 · p. 19

A.7 Additional Details on Baseline Implementation

• FracGen still outperforms fine-tuned baselines trained on the same FracSim data distribution, with Table 4 comparing zero-shot and fine-tuned variants.The matched training protocol uses each baseline’s original default settings; force conditioning differs by model, and text-only prompts cannot express pull direction or magnitude.Source: §A.7 · p. 20, Table 4 · p. 21

A.8 Additional Results

• Ablations indicate that physics-map prediction improves fidelity of bread softening before fracture, while physics-consistency loss improves consistency and accuracy of predicted stress, strain, and damage maps.Figure 11 presents the qualitative ablation comparisons.Source: §A.8 · p. 20, Figure 11 · p. 21

A.9 More Related works

• Prior fracture simulators expose low-level physical controls, whereas FracGen applies comprehensive conditioning to the full stretch-to-fracture progression in a generative model.The targeted kinematic regime includes large stretch, necking, and progressive separation, modeled through damage accumulation rather than plastic flow.Source: §A.9 · pp. 21–22