acceptodds
Under review as a conference paper at ICLR 2027

RealDrop: Predicting Realistic Impact Responses of Dropping

Abstract

Single-image physical reconstruction aims to recover a 3D asset that not only matches appearance but also responds realistically to physical interactions. We focus on object dropping, a particularly revealing setting where impact can trigger diverse and coupled responses, including rebound, deformation, fracture, component motion, liquid release, and splash. However, existing physics-aware reconstruction methods often rely on deformation-dominated physical representations, causing materially complex objects to collapse to overly simple or physically incorrect impact behaviors. We present RealDrop, a framework for predicting realistic dropping responses from a single image. RealDrop constructs part-aware physical assets and introduces a Stress-Responsive Particle System (SRPS) that supports appropriate solid responses from elastic rebound to spatially coherent fracture, together with an SPH-based formulation for liquid dynamics. These representations allow distinct physical behaviors to coexist and interact within heterogeneous objects. We further introduce PhysLab, which turns physical-property inference from appearance-based prediction into consequence-driven verification by executing predicted properties through simulation and iteratively correcting those that produce physically inconsistent outcomes. Extensive experiments demonstrate that RealDrop produces substantially more realistic and expressive impact responses across diverse single-image inputs.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3dgs

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ASa4OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies single-image reconstruction of physically interactive 3D assets, focusing on object responses to dropping. RealDrop combines part-aware reconstruction with a Stress-Responsive Particle System for solid deformation and fracture, and SPH simulation for liquid dynamics. It also introduces PhysLab, a VLM-guided refinement process that proposes expected impact behaviors and iteratively updates physical parameters using rendered simulation feedback. The framework targets coupled responses such as container fracture, component separation, and liquid release. The paper evaluates the system through qualitative comparisons, CLIP similarity, human preference studies, and component ablations, aiming to improve the range and plausibility of simulated impact responses.

Critical strengths
  1. A relevant and expressive system design. Combining part-aware reconstruction, SRPS solid simulation, and SPH liquid simulation addresses heterogeneous impact responses that are difficult to represent with a single material assignment (Sections 3.1-3.3, pp. 3-7). Supporting fracture, component separation, and liquid release within one reconstructed asset is a potentially useful contribution.

  2. Concrete mechanisms and qualitative evidence. The paper specifies energy-based bond failure and finite crack-surface constraints mathematically (Equations (1)-(8), pp. 5-6). The qualitative examples and component ablations illustrate the intended benefits for fragmentation and coupled solid-liquid behavior (Figures 4, 6, 7, and 11, pp. 8-9 and 15). These results motivate further consideration of the system despite the unresolved quantitative validation.

Critical weaknesses
  1. W1: The evaluation protocol and comparison provenance do not yet support a reliable assessment of the claimed realism advantage. Section 4.1 describes pairwise evaluation with 100 participants (p. 7), but UPR denominators and normalization are unspecified. The entries in each of Tables 1-3 sum to 100%; this could reflect normalized pairwise wins and is not inherently erroneous. More directly, Table 3 defines UPR as the full agent's preference rate against each alternative, yet reports 15%, 7%, and a separate 78% for Full Model (p. 9). This ambiguity requires vote counts and aggregation details. Table 1 also transparently includes CLIPsim scores imported from other papers, whose comparability with the authors' 50-object evaluation is not established (p. 7). Finally, the PhysAgent comparison in Figure 5 and Appendix D.1 cites a remote heart-rate estimation paper (pp. 7, 11, and 14), leaving the intended baseline and image source unclear. Together, these issues prevent the quantitative and qualitative comparisons from substantiating the claimed advantage with confidence.

  2. W2: The evidence for PhysLab does not adequately separate physical verification from hypothesis agreement and additional computation. PhysLab adjusts parameters by comparing simulations with an initially predicted behavior hypothesis (Section 3.3, pp. 6-7); the description does not establish whether an incorrect hypothesis can itself be revised. CLIPsim is both feedback and an evaluation metric, including in Figure 10 (p. 15), so its improvement does not independently verify physical accuracy. The paper acknowledges this metric's limitations and provides human evaluation, but W1 limits that complementary evidence. Even after the preference protocol is clarified, preference results would support perceived physical realism, without by themselves establishing that the simulations reproduce the objects' actual physical responses. Claims of physical verification therefore require independent support or a more limited interpretation. Moreover, the comparison with single-pass alternatives in Table 3 does not isolate the benefit of simulation feedback from additional inference and simulation budget (Section 4.4, p. 9). A control using the same initial hypotheses and simulator with matched inference and simulation budgets would help establish this attribution.

  3. W3: Missing implementation and baseline-training details limit reproducibility and assessment of the comparisons. The paper specifies Qwen 3 and at most five PhysLab refinement rounds (pp. 4 and 7), but leaves the exact prompt templates, admissible parameter ranges, and concrete parameter-update procedure unclear. The 24-Layer Transformer in Table 3 also lacks training-data, target, and optimization details (Section 4.4, p. 9), preventing assessment of whether it provides an adequately trained comparison. For SRPS, Equations (1)-(8) describe the failure criterion and crack-surface constraints (pp. 5-6), but do not specify the initial material-graph construction, the calibration procedure for kappa, the stress-averaging and local-peak neighborhood settings, or how initial crack disks and subsequent disk radii and normals are chosen. These details are needed to reproduce the simulator and interpret the inference-strategy comparison.

Questions for authors
  1. What comparisons did participants see, and what are the UPR denominators and aggregation rules in Tables 1-3? Please provide pairwise win/total-vote counts and uncertainty estimates, and reconcile Table 3's caption with its entries (pp. 7 and 9).

  2. Which Table 1 scores were measured under the common 50-object protocol, and what conclusions remain supported by those comparisons? Please also identify the intended PhysAgent baseline and original image source, given the mismatched reference (pp. 7, 11, and 14).

  3. Can PhysLab revise an incorrect initial behavior hypothesis, or only adjust parameters to realize it? What evidence beyond feedback CLIPsim supports physical correctness, and should the claims instead be limited to improved perceived realism where only preference evidence is available (Section 3.3, pp. 6-7; Figure 10, p. 15)?

  4. Does the advantage persist against a control with the same VLM, initial hypotheses, and simulator, and matched inference and simulation budgets, but no feedback-based parameter updates? Please specify the inference and simulation counts used in the comparison (Section 4.4, p. 9).

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

This paper aims to predict how objects respond when dropped, using only a single image. The method is called RealDrop. First, it reconstructs a part-aware 3D Gaussian object using OmniPart and TRELLIS. Qwen3 decides whether each part is solid or hollow and in-paints hidden content, such as liquid in a bottle. For solid, the authors propose SRPS. It uses energy-based bond failure, stress-guided crack direction, local peak selection, and crack-disk constraints, so the fracture produces coherent fragments rather than too many small pieces. For liquids, they use SPH. Then PhysLab uses Qwen3 in a loop. The VLM first hypothesizes drop behavior, then simulates, renders, and feeds frames and CLIPsim back to the VLM to update parameters, for up to 5 rounds. The experiment is on 50 objects from Objaverse and PhysAssets; the metrics are CLIPsim and user preference rate (UPR). RealDrop gets the best number, and ablation shows each component is useful.

Critical strengths
  1. The paper's core motivation is clear. The argument “if the simulator can not express fracture, better parameters are useless” makes sense.
  2. Dropping is a good testbed. One impact can show many different responses: rebound, fracture, liquid. This is harder compared to previous elastic bouncing demos.
  3. SRPS design is reasonable and sound. Compared to the threshold spring-mass, which has the “all broken or not broken” dilemma, peak selection plus a crack disk is a simple way to make fragments coherent.
  4. Qualitative results look amazing. As a VFX experiencer, I think the red wine cases are a great demo to showcase the proposed framework for simulating multiple physical phenomena on heterogeneous materials.
Critical weaknesses

The task itself is ill-posed, but the paper does not discuss. The same bowl from 10 cm to carpet and from 1.5 m to tile have very different results. The image cannot tell the drop height, ground material, or wall thickness. So “predicting” the correct response from a single image is not well defined. The paper never varies these conditions. I worry the method just always breaks fragile-looking objects.

No physical ground truth in evaluation. No real drop video, no fragment statistics, no comparison to real outcome. UPR only measures “look realistic”. Humans and VLMs may share the same stereotype, like “ceramic must break”. So UPR can not separate physically correct from dramatic.

UPR (user preference rate) protocol is unclear. The text says pairwise, but the numbers in Tables 1, 2, and 3 all sum to 100% and look like multi-choice shares. Table 3 caption says “how often the full model is preferred”, which also does not match. UPR is the main evidence of the paper, so this needs to be clear.

Novelty of SRPS vs. graphics fracture work is not discussed. Brittle fracture has been studied for a long time in graphics (FEM fracture, peridynamics, CD-MPM, etc.). The paper only compares with naive threshold spring-mass and plain MPM. Also, crack initiation details (when the first disk forms, how r_s is set, how the normal is decided) are missing. No sensitivity study for η, r_s, or particle resolution.

“24-Layer Transformer” baseline is not explained. No training data, no architecture details. So the claim “advantage comes from closed-loop, not model capacity” is not supported.

Questions for authors
  • If you change the drop height and ground material (for example, 10 cm vs. 1.5 m, or carpet vs. tile), does RealDrop give different results for the same object?
  • Is there any case where a fragile-looking object correctly does not break? How often is the initial VLM hypothesis wrong, and can PhysLab correct the hypothesis itself, or only the parameters to match it?
  • Can you qualitatively compare with real drop videos with real outcomes?
  • Please explain the UPR protocol exactly. Is it pairwise or multi-choice? How many videos does each participant see, and is the order randomized?
  • How does SRPS compare to existing fracture methods such as CD-MPM and peridynamics? What are the implementation details?
  • What is the 24-Layer Transformer baseline? Please clarify its training data, input and output, and architecture details.
Reviewer_dmq2OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

RealDrop constructs simulation-ready 3D assets from single images to generate realistic dropping responses. It combines part-aware Gaussian reconstruction, a Stress-Responsive Particle System (SRPS) for solid deformation and coherent fracture, and SPH-based liquid simulation. PhysLab uses a VLM to hypothesize expected impact behavior and iteratively revise physical parameters using rendered simulation feedback and CLIP similarity. Evaluation uses 50 objects from Objaverse and PhysAssets, frame-to-prompt CLIP similarity, and a reported 100-participant preference study, alongside component ablations and qualitative examples. The central claim is that the integrated framework produces more realistic coupled responses, including fracture, component separation, and liquid release.

Critical strengths
  1. A useful focus on coupled responses during freefall impact. The paper targets situations in which different components must respond differently to the same impact, such as a container fracturing while its contents separate or flow. This is a meaningful application focus for physical asset generation. I give this limited positive weight as motivation: part-wise physical modeling and simulation already appear in work such as Phys4DGen, while GaussianFluent addresses mixed-material fracture.
Critical weaknesses

1. Additional evaluation details and clarification of the validation scope would help interpret the results.

Section 4.1 outlines the CLIPsim calculation and a 100-participant pairwise preference study. It would be helpful to provide examples of the evaluation motion prompts and explain how they were constructed, including whether they describe only the action or also the expected outcome, and how they relate to PhysLab's behavioral hypotheses.

The human evaluation would also benefit from a description of the comparison pairs, judgment counts, participant assignment, information shown to participants, and presentation controls. In particular, please clarify the UPR denominator and aggregation rule. Table 3's caption describes preference for the full agent over each alternative, while the rows report 15%, 7%, and 78%; the values in each of Tables 1–3 also sum to 100%. Different comparison sets can reasonably yield different full-model scores, but an explicit definition would help readers understand whether the entries are pairwise win rates or normalized shares of preferences. These clarifications would make the evaluation easier to interpret and reproduce.

The scope of validation could also be clarified. PhysLab adjusts properties to match expected behavior in a simulated interaction, but it remains unclear whether the resulting asset responds plausibly to different drop heights, initial orientations, and contact conditions without further parameter refinement. The in-the-wild and stretching/bending examples in Appendices D.2–D.3 are useful, but do not clearly establish reuse of the same inferred properties across conditions. Clarifying this would help distinguish reusable physical assets from configurations tuned to a particular interaction. The concern is generalization within the intended setting, rather than the choice to focus on dropping.

2. The manuscript would benefit from improved formatting and reporting consistency.

Figure 5 appears before Figure 4, and extensive text wrapping around figures and tables on pp. 7, 9, and 15 creates narrow text blocks and interrupts the reading flow. A careful layout and cross-reference check would improve the presentation.

Additionally, the PhysAgent citation associated with Figure 5 and Appendix D.1 points to a heart-rate estimation paper. The intended reference may instead be PhysAgent: Automating Physics-Based 4D Synthesis via Trajectory-Grounded Multi-Agent Feedback, by Lv et al. (2026). Please check the citation and identify the source of the displayed comparison result. These appear to be correctable presentation and reporting issues.

Questions for authors
  1. Evaluation prompts: What exact motion prompts were used, and how were they constructed? Do they specify expected physical outcomes, and what is their relationship to PhysLab's VLM-generated hypotheses?
  2. Human evaluation and UPR: Could the authors describe the comparison and participant protocol and explain how preferences were aggregated into Tables 1–3? In particular, how should the Table 3 caption and its entries be read together? Please also clarify Appendix A's statement in light of the reported 100-participant study.
  3. Reuse across conditions: Are an asset's inferred properties held fixed when drop height, orientation, or contact conditions change, or is PhysLab rerun? What existing results demonstrate reuse without refinement? If this has not been tested, a small controlled comparison within the dropping setting would help clarify the scope of the method.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.