acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

We introduce FracGen, a fracture-aware video generation model that produces plausible, controllable fracture dynamics from a single image of an intact object, conditioned on physics signals. To train FracGen, we build FracSim, a fracture-aware simulation framework that augments material point method (MPM) simulation with a continuum damage model, producing paired fracture videos and dense, pixel-aligned physical fields at no additional cost beyond standard rendering. FracGen leverages these maps in two ways: it is trained to jointly predict them alongside RGB video, encouraging the model to capture physical state rather than surface appearance; and it is supervised with physics-informed losses that encourage consistency among the predicted maps. As a result, FracGen captures distinct material-specific fracture behavior without expensive test-time simulation or per-scene tuning, while offering fine-grained control over where an object tears, how fast the crack propagates, and how much deformation precedes failure. We further introduce a benchmark for evaluating the physical plausibility of generated fracture video, and show through extensive experiments that FracGen outperforms existing video generation baselines in both physical and visual fidelity.

Primary area:
generative models
Keywords:
video generation models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_euxgOpenReview ↗
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
  1. 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.

  2. 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.

  3. 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
  1. 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.

  2. 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.

  3. 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
  1. 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?

  2. 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?

  3. 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 sigmoid⁡(z^damage)\operatorname{sigmoid}(\hat z^{\mathrm{damage}}) 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?

  4. 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
AuthorsOpenReview ↗

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.

  • Quantitative, held-out objects. All numbers in Tables 1 and 2 are computed on four objects never seen in training, across 2,808 test samples covering the full range of loading, material and fracture conditions. On these unseen objects, FracGen reaches FVD 266.17 and SAC 0.946.
  • Qualitative, in-the-wild images. The objects in Fig. 4–6 come from a further set of ten single images with no 3D reconstruction or simulation (Fig. 8). On these, FracGen produces controllable stretching and tearing under varying force directions, material properties and fracture properties. Results are best viewed on the our project website (https://iclr11751.github.io/).

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.

AuthorsOpenReview ↗

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.

MethodApplied loadingMaterial propertiesFracture propertiesPredicts physics maps
CogVideoX-5B-I2V-FTForce-aware text caption✗✗✗
ForcePrompting-FTTwo-grip Gaussian blobs✗✗✗
PhyCo-FTTwo force maps✓✓✗
PhysCtrl-FTTwo 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.

AuthorsOpenReview ↗

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.
Reviewer_hXSUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces FracGen for generating strectching and tearing videos given a single image with material properties as condition. The authors augment the PhysGaussian MPM with stetch-driven damage model and produces RGB video supervision with motion, strain etc. The authors use pretrained WAN to jointly predict RGB images and these physical maps with physics-consistency loss. Experiments on held-out objects demonstrate the controllability and quality of proposed method.

Critical strengths
  • The authors focused on an interesting and less studied problem in physics-aware video generation.
  • The authors proposed a reasonable pipeline and training scheme and demonstrated its effectivenes for fracture-aware video generation.
  • Good ablation study that demonstrates the usefulness of the components.
Critical weaknesses
  • The authors should compare with standard Image-to-video models, like Wan-I2V.
  • The paper lacks results for diversity and real-world images, the results look synthetic and artificial, the videos also contain noticeable artifacts of blurriness possibly caused by Gaussian rendering.
  • "We do not compare against PhysCtrl (Wang et al., 2025), as it encodes force as a single force vector, which cannot represent our dual opposing forces." Compare to PhysCtrl should be easy since opposing forces can be represented with multiple force vector tokens.
Questions for authors

What are the value of λconstλ_{const} and the exact definition of w(t)w(t)?

How many optimization steps are used in each training stage?

3 comments
AuthorsOpenReview ↗

Response to Reviewer hXSU

We thank the reviewer for the constructive feedback, and for acknowledging that our work focuses on solving an interesting and less studied problem. We are also grateful for positive assessment of our FracGen as a reasonable and effective pipeline and training scheme for fracture-aware video generation, with good ablation study to demonstrate the usefulness of the components. We address the reviewer's concerns below.

1/ Comparison with Wan-I2V and PhysCtrl. We thank the reviewer for this suggestion. For PhysCtrl, we follow reviewer's suggestion to use multiple force tokens to represent for opposing forces. And we fine-tuned PhysCtrl on the same simulated data, using the same fracture and material properties by adding additional tokens for them, so that PhysCtrl-FT receives the same conditions as FracGen. For Wan-I2V, we compare with 14B variant checkpoints. As shown in the following table, FracGen outperforms both PhysCtrl-FT and Wan-I2V on all metrics under equal conditioning. Qualitative comparisons are available on our project website (https://iclr11751.github.io/). Thanks again for suggesting this as a strong baseline and we incorporated in our revised manuscript.

ModelFVD ↓LPIPS ↓PSNR ↑SSIM ↑SAC ↑
Wan2.2-I2V-14B1196.10.32314.370.7980.401
PhysCtrl-FT582.60.17219.260.8510.793
FracGen (Ours)266.170.150221.190.8600.946

2/ Diversity and real-world images. We thank the reviewer for this question. We would like to clarify that the paper already includes diverse results beyond the simulated objects. Fig. 4, Fig. 5 and Fig. 6 show FracGen on additional single images (listed in Fig. 8) that have no 3D reconstruction or simulation and that differ from the training objects in geometry and appearance. More results on these objects, under varying forces, material properties and fracture properties, are available on our project website (https://iclr11751.github.io/).

The reviewer is right that the visual artifacts come from the Gaussian Splatting renders used during simulation, which carry over into our training data. Gaussian Splatting has seen several recent improvements, and these can be incorporated into our data generation to remove such artifacts. We therefore see this as a limitation of the current renderer and not of our approach: the main contribution of this paper is how to extract physical maps from simulation that are useful for training a generative model of fracture. We believe this recipe will continue to hold as Gaussian Splatting and other radiance-field methods improve. To the best of our knowledge, our FracGen is the first work to show that a video generator can be repurposed for controllable stretching and tearing.

Real-world images. We have added results on real photographs for both low-stretch and high-stretch objects. Please see point 2 of the general response above.

3/ Value of λ_const, definition of w(t), and number of optimization steps in each training stage We thank the reviewer for pointing out these missing details. We set λ_const = 0.1 during training Stage 3 as a weight for the regularization term. The weight w(t) is a decaying function of the noise level. The constitutive loss is evaluated on the predicted clean latent, obtained in one step from the noisy latent and the predicted velocity. At high noise levels this estimate is a poor approximation of the final sample, so enforcing a stress–strain–damage relation on it would mostly penalize estimation error. The weighting switches the regularizer off in that regime and applies it fully near clean latents, where the estimate is reliable and the physical maps are well defined.

Using the convention of Eq. (1), where t ∈ [0, 1] and t = 1 corresponds to the clean latent, the weight is defined as

w(t) = t^γ, with γ = 2.

In our implementation, with discrete timesteps τ ∈ [0, T], T = 1000 and τ = 0 as the clean latent, this is w = (1 − τ/T)^γ. The weight is 1 for clean latents and decays quadratically to 0 at pure noise. We have added these details in our revised manuscript.

For multi-stage training, each stage is trained for 7,000 steps, and each one is initialized from the checkpoint of the previous stage: Stage 1 (RGB prediction with physics conditioning), then Stage 2 (joint prediction of RGB and FracPhys Maps), then Stage 3 (full objective with the constitutive loss). The final model has therefore been trained for 21,000 steps in total. We have added this to the training details in Appendix A.4.

AuthorsOpenReview ↗

Follow-up on rebuttal response

Dear Reviewer hXSU,

Thank you for your thoughtful review and helpful suggestions. We have posted our response, which includes the additional comparisons with PhysCtrl and Wan-I2V that you suggested, along with a general response and a revised manuscript. With the discussion period coming to an end soon, we would be very grateful if you could let us know whether our response addresses your concerns. We are happy to clarify anything further.

Best,

The Authors

AuthorsOpenReview ↗

Follow-up on our rebuttal

We thank the reviewer once again for the constructive feedback and we have posted our rebuttal response to address reviewer's concern. As we are approaching the end of the discussion period, we would really appreciate if the reviewer could take a look at our response to see if it well addresses your concern and we are happy to clarify anything further.

Reviewer_tMooOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces FracGen, a conditional video generation model for object stretching and tearing. The model takes an initial image, prescribed loading, material parameters, and fracture parameters, and jointly generates RGB video and dynamic physical maps. Training data are produced by FracSim, which augments Gaussian–MPM simulation with a stretch-driven damage model. The method adapts a pretrained video model using LoRA and introduces a latent consistency loss coupling stress, strain, and damage. Experiments on held-out objects from the simulated dataset show improvements over fine-tuned video generation baselines.

Critical strengths
  1. Jointly generating fracture videos and aligned physical fields is an interesting approach. It provides additional supervision and makes the predicted physical evolution available for inspection, beyond judging RGB appearance alone.

  2. The paper includes comparisons with baselines fine-tuned on the same dataset and an internal ablation of physical-map prediction and the consistency loss. The consistency loss improves both FVD and several physical-map metrics.

Critical weaknesses
  1. The conditioning inputs are not individually ablated.

FracGen receives detailed loading, material, and fracture parameters, but Table 2 only ablates physical-map prediction and the consistency loss. The necessity of each condition, particularly the discrete fracture category, remains unclear. Baselines also lack equivalent conditioning, while the RGB-only Stage 1 already achieves most of the improvement over them.

  1. Physical realism and generalization remain insufficiently validated.

The benchmark uses the same simulator as training and only four held-out objects. The simplified damage model lacks independent physical validation.

Questions for authors
  1. Could you provide ablations of the individual conditioning inputs to clarify their contributions and possible redundancy?
  2. Could you provide more video examples across diverse objects and conditions, including failure cases?
  3. Beyond held-out objects generated with the same simulator, have you evaluated on an out-of-distribution benchmark, such as unseen material–parameter combinations or real-world fracture videos?
5 comments
AuthorsOpenReview ↗

Response to Reviewers tMoo (1/2)

We thank the reviewer for the constructive feedback and for recognizing FracGen’s joint generation of fracture videos and aligned physical fields as an interesting approach, and for noting that it adds supervision and makes the predicted physical evolution available for inspection beyond RGB appearance alone. We also appreciate the reviewer acknowledging our comparisons with baselines fine-tuned on the same dataset, and our ablation of physical-map prediction and the consistency loss, which shows that the consistency loss improves both FVD and several physical-map metrics. We address the reviewer’s concerns below.

1/ Ablation of the conditioning inputs. We thank the reviewer for this question. We ran the requested ablation and report it in the general response above, together with our analysis. Please see point 1 of the general response above.

2/ Why the benchmark uses the same simulator. Our metrics for physical alignment need dense ground truth for stress, strain, damage and flow, and known material parameters and grip motion. Only simulation provides these, which is why the quantitative benchmark is built on held-out objects from FracSim.

3/ More video examples and failure cases. We thank the reviewer for this question. Our project website (https://iclr11751.github.io/) already provides extensive video results showing that FracGen generates plausible stretching and tearing across diverse objects and conditions, including different forces, material properties and fracture properties. We have also added new results there under the Rebuttal section, including results of FracGen when tested on real-world image

Failure cases. One failure mode appears on real-world images. As shown in the real-image videos in the Rebuttal section of our project website (https://iclr11751.github.io/), FracGen still generates plausible stretching and tearing of the object, but the camera viewpoint can drift slightly during the video. We attribute this to a gap between training and test data: our training videos contain isolated foreground objects without a surrounding scene, and provide no explicit camera supervision or control, so the model has not learned to keep the background of a real photograph fixed. Addressing this, for instance by adding scene backgrounds or camera conditioning to the training data, is left for future work.

AuthorsOpenReview ↗

Response to Reviewers tMoo (2/2)

4/ Out-of-distribution benchmark and real-world fracture video.

a) Evaluation of unseen material-parameter combinations. Thank you for the suggestion. We ran a new evaluation in which FracGen is given material and fracture parameters combination that never appear in training, on the four held-out objects, so that both the object and the parameter values are unseen. We designed three sets, each targeting one type of shift:

  • Set A, interpolated stiffness: Young's modulus values that fall between the training values.
  • Set B, extrapolated stiffness: values below the smallest and above the largest training value, by a factor of 1.5 to 2.
  • Set C, unseen fracture thresholds: onset and critical stretch pairs not used in training.

We summarize these value sets as follows:

SetBread-like (croissant, muffin)Rubber-like (two held-out toys)Clips
Training gridE ∈ {2k, 6k, 10k} Pa; (λ_on, λ_cr) ∈ {(1.5, 2.3), (2.0, 3.0)}E ∈ {20k, 60k, 100k} Pa; (λ_on, λ_cr) ∈ {(2.6, 3.0), (3.0, 3.3)}–
A: interpolated stiffnessE ∈ {4k, 8k}E ∈ {40k, 80k}16
B: extrapolated stiffnessE ∈ {1k, 15k}E ∈ {10k, 150k}16
C: unseen thresholds(λ_on, λ_cr) ∈ {(1.7, 2.6), (2.4, 3.2)}(λ_on, λ_cr) ∈ {(2.0, 2.6), (2.8, 3.15)}16

Results. The reference row is FracGen on the full held-out test set with training-grid parameters (Table 2). We omit FVD for the unseen sets because it is unreliable at this sample size.

SetLPIPS ↓PSNR ↑SSIM ↑SAC ↑DPA ↑Map MSE ↓R²_cons ↑
Reference: seen parameters (Table 2)0.15021.190.8600.9460.5360.01750.902
A: interpolated stiffness0.16020.650.8560.9480.4550.01540.938
B: extrapolated stiffness0.15520.860.8600.9400.4910.01540.859
C: unseen thresholds0.15520.710.8600.9600.5330.01470.922

Performance is largely maintained on the unseen sets. Video quality (LPIPS, PSNR, SSIM), pull-axis alignment (SAC) and physics-map error remain comparable to the reference. The largest changes are in damage timing under interpolated stiffness (DPA 0.455 against 0.536) and in constitutive consistency under extrapolated stiffness (R²_cons 0.859 against 0.902), which shows that the predicted physics is less reliable once stiffness leaves the training values. Overall, the model is not limited to the parameter values it was trained on.

b) Testing on real-world fracture videos. We have added qualitative results on real-world fracture videos. Please see the general response above for these results and our discussion.

AuthorsOpenReview ↗

Follow-up on our rebuttal response

Dear Reviewer tMoo,

Thank you again for your constructive feedback. We have posted our response to address your concerns, together with a general response and a revised manuscript. As the discussion period is approach to the final stage, we would be grateful if you could take a look and let us know whether our response addresses your concerns. We are happy to provide any further clarification.

Best,

The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to Reviewers

We thank all reviewers for their constructive feedback and positive assessment of our work. Specifically, the reviewers found that FracGen addresses an interesting and less studied problem in physics-aware video generation (hXSU), that jointly generating fracture videos and aligned physical fields is an interesting approach that makes the predicted physical evolution available for inspection (tMoo), and that it is the first framework to do so from a single image, with FracSim providing controllable fracture simulation and dense physical supervision (euxg). The reviewers also appreciated our comparisons with fine-tuned baselines and our ablation studies (tMoo, hXSU). We address the common concerns below and respond to each reviewer individually. We also made changes in the revised manuscript, which are marked in blue.

1/ Ablation Conditioning Input to FracGen (Reviewer tMoo, and euxg).

We retrained our model with reduced sets of conditions under an identical training setup: a) force only, b) force and material properties, and c) force and fracture properties.

Ablating Input ConditionsFVD ↓LPIPS ↓PSNR ↑SAC ↑
a) Force only290.20.16020.660.940
b) Force, Material Properties273.40.15720.730.932
c) Force, Fracture Properties282.90.15920.810.925
FracGen, Full266.170.15021.190.946

As shown, we demonstrate that although combination of material and fracture properties can slightly improve video quality, the meaning of these conditions is not to improve video quality but rather allow the user to perform controllable stretching and tearing experiment. These are shown in controllable experiment with controllable stiffness and fracture timing (Fig. 1, Fig. 7). Additionally, these properties are important for the prediction of strain, stress and damage map. Without these inputs, the model cannot be asked to predict FracPhys Maps for a specified material, and it offers no control over when or whether the object tears. We have added this ablation study in the revised manuscript.

Additionally, these aggregate metrics do not fully capture physical plausibility, so we also encourage the reviewers to view the ablation videos on our project website (https://iclr11751.github.io/) and the qualitative comparison in Fig. 12. They visually show how joint prediction improves the tearing behavior, for example softening before the break, and how the constitutive loss keeps the predicted stress, strain and damage maps consistent with one another. We have included discussion on the conditioning ablation results in the revised manuscript.

2/ Testing on real-world images (Reviewers tMoo and hXSU). We have added results on real-world images to the revised paper (Fig. 6). Specifically, we captured real objects placed on a table and gave each photograph to FracGen as the input image, together with the force, material and fracture conditions. As shown, FracGen produces plausible stretching and tearing on these real photographs for both low-stretch and high-stretch objects. The results are best viewed as videos in the Rebuttal section of our project website (https://iclr11751.github.io/).

AuthorsOpenReview ↗

Follow-up on our rebuttal response

Dear Reviewers,

Thank you once again for your constructive feedback on our paper. We would like to kindly follow up on our rebuttal response, which we have answered questions from all reviewers. As we approach the end of the discussion period, we would greatly appreciate it if you could take a look at our responses when you have a chance.

Please let us know if you have any further questions or concerns. We would be happy to provide additional clarification. Thank you for your time and consideration!

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.