- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper presents FracGen, a video generation framework that jointly predicts RGB fracture videos and flow, strain, stress, and damage maps from a single object image, conditioned on applied loading, material properties, and fracture properties. To provide training supervision, the authors introduce FracSim, which augments Gaussian–MPM simulation with an irreversible damage model driven by maximum principal stretch and renders pixel-aligned physical fields alongside RGB videos. FracGen adapts a pretrained Wan 2.1 video generator using LoRA and incorporates a latent constitutive regularizer inspired by constitutive modeling and damage-induced stress degradation. Experiments evaluate video quality and physical alignment on a simulation-based benchmark, complemented by ablation studies and qualitative results on unseen object images.
Critical strengths
-
The first framework for jointly generating material-dependent tearing videos and physical fields from a single image. FracGen is the first video generation framework to jointly predict material-dependent tearing and the associated physical fields from a single image and specified physical conditions. This work extends physics-conditioned video generation to the full deformation-to-fracture progression and introduces a corresponding task formulation and evaluation benchmark.
-
FracSim provides controllable fracture simulation and dense physical supervision. The authors incorporate an irreversible continuum damage model into Gaussian–MPM to simulate the progression from tensile deformation to material failure, while simultaneously rendering pixel-aligned RGB videos and flow, strain, stress, and damage maps. By varying applied loading, material properties, and fracture properties, FracSim produces paired training data with diverse fracture behaviors and rich physical supervision.
-
FracGen combines physics conditioning with joint prediction and achieves state-of-the-art performance on the reported benchmark. FracGen conditions a pretrained video generator on applied loading, material properties, and fracture properties, and jointly predicts RGB fracture videos and dynamic FracPhys Maps to learn relationships between appearance evolution and physical state. A latent constitutive regularizer inspired by constitutive modeling and damage-induced stress degradation further improves prediction. FracGen outperforms the compared methods across FVD, LPIPS, PSNR, and SAC, reducing FVD from the best baseline’s 832.12 to 266.17. Ablation studies support the benefits of the constitutive loss for video quality and overall physics-map accuracy, while qualitative results demonstrate better appearance preservation and controllable tearing behavior.
Critical weaknesses
-
Unequal conditioning information confounds baseline comparisons. Appendix A.7 states that ForcePrompting has “no conditions on other channels such as Young’s modulus, Poisson’s ratio, or fracture thresholds,” while CogVideoX uses “identical captions … for opposing pull directions of the same object.” Equivalent material and fracture conditioning for PhyCo is also unclear. Thus, the reported gains may partly reflect access to additional conditions rather than the proposed joint prediction and constitutive loss.
-
Training-object diversity is limited, and evidence of cross-object generalization remains insufficient. The 4,212 training samples are derived from only six objects—three low-stretch and three high-stretch objects—with diversity primarily arising from variations in physical conditions.
-
The physical justification for the latent constitutive regularizer is insufficient. Equations (8)–(10) apply forms inspired by linear elasticity and damage-induced stress degradation to features encoded by a frozen VAE, but nonlinear encoding generally does not preserve linear and multiplicative relationships in physical space. In particular, applying a sigmoid directly to the damage latent does not guarantee that the gate varies monotonically with physical damage; likewise, driving the stress target toward zero latent does not imply zero physical stress. Although the authors acknowledge that this is an approximate regularizer, the ablation only establishes the empirical benefit of the complete loss, without distinguishing the contribution of its physical structure from that of generic cross-modal consistency regularization.
Questions for authors
-
Could the authors clarify the conditions provided to each baseline and provide comparisons with equivalent inputs, or controlled ablations that isolate the benefits of the additional material and fracture conditions?
-
Can FracSim and FracGen be integrated into broader physics simulation and video generation frameworks beyond stretch-to-tear fracture? Which components are transferable, and what modifications would be needed to support other physical interactions or failure modes?
-
Could the authors explain or validate how the latent constitutive regularizer in Eqs. (8)–(10) reflects stress–strain relationships and damage-induced stress degradation in physical space, particularly the correspondence between and physical damage? Could evaluations of physical relationships after decoding, together with comparisons against a capacity-matched generic cross-modal regularizer, distinguish the contribution of the physical structure from general regularization benefits?
-
Could the authors provide quantitative evaluations on a broader set of objects that differ more substantially from the training objects in geometry, appearance, and category?
9 comments
Response to Reviewer euxg (2/4)
2/ Clarification of conditions provided to each baseline. We thank the reviewer for this thoughtful question. We clarify by summarizing each method's inputs in the table below.
| Method | Applied loading | Material properties | Fracture properties | Predicts physics maps |
|---|---|---|---|---|
| CogVideoX-5B-I2V-FT | Force-aware text caption | ✗ | ✗ | ✗ |
| ForcePrompting-FT | Two-grip Gaussian blobs | ✗ | ✗ | ✗ |
| PhyCo-FT | Two force maps | ✓ | ✓ | ✗ |
| PhysCtrl-FT | Two force tokens | ✓ | ✓ | ✗ |
| FracGen (ours) | Two-grip Gaussian blobs | ✓ | ✓ | ✓ |
For CogVideoX-5B-I2V-FT, the text prompt is the only input through which a force condition can be specified, so we describe the applied force in the caption. The prompt we used is given in Appendix A.7.
For PhyCo-FT, we repurposed its pixel-aligned physical property channels to carry our material and fracture properties, so it receives the same three groups of conditions as FracGen. We additionally include a PhysCtrl baseline fine-tuned on our setting, as suggested by Reviewer hXSU. For PhysCtrl-FT, we added tokens representing the material and fracture properties, so it also receives the same conditions as FracGen.
ForcePrompting-FT and CogVideoX-5B-I2V-FT, in contrast, provide no native interface for material or fracture parameters. Adding one would require substantial methodological and architectural changes to these methods, which would no longer be a faithful comparison with the original designs and is beyond the scope of our baseline study. For these two methods, the matched-input comparison is instead provided by our force-only ablation in point 1 of the general response. We have revised Appendix A.7 to state this explicitly.
In summary, ForcePrompting-FT, PhyCo-FT and PhysCtrl-FT are all physics-informed generation methods, but none of them was designed for our task: controllable stretching and tearing, where large deformation precedes material failure. FracGen targets this behavior directly through three components: a damage model incorporated into the simulation that synthesizes our training data, joint prediction of RGB video and physical fields, and a physics-consistency loss. As a result, FracGen outperforms these methods both quantitatively and qualitatively (Table 1, main paper), including both PhyCo-FT and PhysCtrl-FT, which is given equivalent conditions.
Beyond video quality on stretching and tearing, FracGen offers two capabilities the baselines lack. It provides control through several physical inputs (applied loading, material properties and fracture properties), and it predicts flow, strain, stress and damage maps, which make the generated dynamics available for inspection.
3/ Reported gains may partly reflect access to additional conditions. We thank the reviewer for raising this. Two results address it. First, PhyCo-FT and PhysCtrl-FT receive the same conditions as FracGen and still fall well behind (We have updated Tab. 1 in our revised manuscript). Second, a version of our model that receives the force condition only still reaches FVD 290.2, against 832.12 for PhyCo-FT (general response, point 1). The margin is therefore not explained by access to additional conditions.
Response to Reviewer euxg (3/4)
4/ Intergration of FracSim and FracGen. We thank the reviewer for this insightful question. Our work targets stretching and tearing, and several mechanisms are designed specifically for it. However, the overall recipe is general: (i) add the mechanism governing the target physics to the simulator, (ii) render its internal state as pixel-aligned maps, and (iii) train the generator to predict these maps jointly with RGB under the corresponding conditions. We view FracSim and FracGen as one worked instance of this recipe, and we hope future work can benefit from this.
We provide our discussion below what transfers directly from what must be redesigned.
FracSim.
- Transferable. The physics-map renderer (Eq. 6) composites any per-particle quantity through the Gaussian rasterizer, so it is independent of the physics being simulated. Flow, strain and stress are available from the MPM state under any constitutive model, and MPM itself already supports other materials such as plastic, granular and fluid-like media.
- Needs modification. The broader mechanism must match the target physics, in the same way that we adopted continuum damage to introduce a damage variable, control fracture timing and render a damage map.
FracGen.
- Transferable. Joint latent construction (Eq. 7) encodes each map with the frozen VAE and concatenates it along the height axis, so adding or replacing a physical field requires no architectural change. The conditioning interface also extends naturally: spatial signals are rendered as control videos and scalar parameters are broadcast as channels.
- Needs modification. New physics requires deciding how to represent its conditions and feed them to the generator, for example an impactor's contact point and velocity for collisions, or a yield stress for plasticity. Our latent constitutive regularizer (Eqs. 8–10) encodes linear elasticity with damage-induced degradation and this could be replaced by the relation governing the new physics. Our SAC and DPA metrics are specific to tensile loading, so new evaluation metrics would also be needed.
Response to Reviewer euxg (1/4)
We thank the reviewer for the constructive feedback and for recognizing FracGen as the first framework to jointly generate tearing videos and physical fields from a single image. We also appreciate the reviewer's assessment that FracSim provides controllable fracture simulation with dense physical supervision, and that FracGen achieves state-of-the-art performance on the reported benchmark. We address the reviewer's concerns below.
1/ Training-object diversity is limited, and evidence of cross-object generalization remains insufficient. We thank the reviewer for raising this. The small training-object set is deliberate. The question we pose in the introduction is whether a video generator can learn how objects stretch and tear from limited data and lightweight fine-tuning. Under limited computing resources, we therefore kept the object set minimal, which also makes the evaluation a test of generalization: the model sees only six objects (three low-stretch, three high-stretch) and must transfer to objects it has never seen. Interestingly, even with this small set, FracGen generalizes to unseen objects with new geometry and appearance (Fig. 4, Fig. 5 and Fig. 6). More results can be seen in our project website (https://iclr11751.github.io/). Note that results shown in our project website are not seen by the model during training.
Diversity comes from physical conditions, not object count. Each object is simulated under different combinations of tearing axis, grip offset, pulling speed, Young's modulus, Poisson's ratio and stretch thresholds (Table 3), giving 4,212 training samples. The model is thus trained to map physical conditions to deformation and tearing behavior, while object appearance and geometry are handled by the pretrained video prior, which LoRA fine-tuning preserves.
Evidence of cross-object generalization.
Scope. Our objects cover two material classes, bread-like and rubber-like, and we do not claim generalization to categories outside them. Extending the object set and material classes requires only additional simulation, which FracSim supports without changes, and we leave this for future work.