Pith. sign in

REVIEW 3 major objections 6 minor 2 cited by

This paper argues that interaction-ready 3D reconstruction can be driven by counterfactual consistency: removing objects and restoring them should recover the original scene, and enforcing that match yields appearance, spatial, and physical

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review

2026-08-04 04:54 UTC pith:D7DKQ6LU

load-bearing objection The abstract sells a counterfactual alignment framework that doesn't exist in the body; what's actually there is a decent but under-validated pipeline paper under a mismatched title. the 3 major comments →

arxiv 2605.30239 v2 pith:D7DKQ6LU submitted 2026-05-28 cs.CV

CA-World: Multi-Object Counterfactual Alignment for Efficient Interactive-Ready Reconstruction

Xin Dong , Weijian Deng , Lihan Zhang , Tianru Dai , Wenfeng Deng , Yansong Tang This is my paper
classification cs.CV
keywords interaction-ready reconstructioncounterfactual alignment3D reconstructionphysics simulationgenerative 3D priorsobject-scene alignmentmulti-object interactionmaterial point method
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that a reconstructed 3D world is only truly ready for interaction if it anticipates change—objects can be moved, removed, and collided with. To build such worlds efficiently, it frames foreground-background decoupling as a visual intervention: removing objects and inpainting the background is a counterfactual edit, and reintegrating the completed objects is the inverse edit. Counterfactual consistency then says that reversing the intervention should return to the observed scene, which yields three supervision signals: appearance, spatial, and physical consistency. This lets each object be checked against the observed scene locally, avoiding joint optimization of all object states. The paper also presents a training-free implementation of this idea and tests it on real multi-object scenes.

Core claim

On the paper's own terms, the central claim is that interaction-ready reconstruction reduces to counterfactual alignment learning. The authors define object removal and background inpainting as a counterfactual intervention, object completion as counterfactual generation, and scene reintegration as the inverse intervention. The requirement that the reintegrated scene match the original in appearance, spatial layout, and physical plausibility provides direct supervision without ground-truth object states. Locality of object-level interventions means each counterfactual state is constrained by the observed scene rather than by jointly optimizing all objects, cutting cost and error accumulation

What carries the argument

The central mechanism is counterfactual consistency over a decoupling-reintegration cycle. Foreground-background separation is treated as a visual intervention; separate object generation and background inpainting are treated as counterfactual generation; and scene reintegration is treated as the inverse intervention. The paper claims that enforcing equality between reintegrated and original scenes in appearance, spatial relationship, and physical plausibility is a sufficient supervision signal, and that the locality of object-level interventions lets each object be aligned independently using the observed scene, without joint optimization. In the implementation, this appears as three post-p

Load-bearing premise

The load-bearing premise is that removing and reintegrating an object is an invertible intervention—that any mismatch between the reintegrated scene and the original genuinely reflects incorrect completion or alignment, rather than generative ambiguity, lighting changes, or inpainting artifacts.

What would settle it

Set up a scene with a known object, remove the object, inpaint the background, and reintegrate the exact original object at its exact pose. If the three consistency losses (appearance, spatial, physical) do not drop near zero—for example because the background inpainting changes lighting or shadows—then counterfactual consistency cannot serve as a faithful supervision signal for interaction readiness.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If counterfactual consistency is a valid supervisory signal, interaction-ready 3D reconstruction can be trained from ordinary multi-view images, since the original scene serves as its own ground truth.
  • Locality of object-level interventions implies reconstruction cost scales with the number of objects rather than jointly with all object states, reducing computational load and error accumulation in cluttered scenes.
  • The three consistency objectives (appearance, spatial, physical) provide a concrete checklist for whether a reconstructed scene will behave correctly under interaction—not merely render well.
  • The training-free implementation can run on consumer hardware and support real-time multi-object physical simulation, making interactive reconstruction practical outside research labs.
  • Clean object-scene separation plus completion enables dynamic scene editing, such as removing, replacing, or re-simulating individual objects.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The abstract promises counterfactual alignment learning, but the full-text implementation is a training-free pipeline that uses a frozen generative prior and optimization-based post-processing; a direct test of the counterfactual claim would be to train the completion model itself with the three consistency losses and measure improvement.
  • The counterfactual framing naturally suggests an automatic evaluation protocol: perform an intervention (remove and reintegrate an object), then measure the three consistencies; scenes that score well should be the ones that simulate correctly. This could become a standard benchmark for interaction-ready reconstruction.
  • The locality argument implies that the method's benefit grows with scene clutter; in sparse scenes the counterfactual supervision may provide little signal, so the approach may be best suited for dense multi-object environments.
  • If the invertibility assumption is violated (e.g., shadows, specular highlights, or inpainting ambiguity), the consistency loss could misattribute generative ambiguity to causal error; extending the framework to model these effects is a natural next step.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The abstract of arXiv:2605.30239 introduces CA-World, an 'efficient framework that integrates counterfactual alignment learning into a decoupling-reintegration reconstruction pipeline,' with three alignment objectives (appearance, spatial, physical) that enforce counterfactual consistency by requiring that reversing an intervention recovers the factual world. The full text, however, is a different paper titled 'SAM3D-Phys: Towards Multi-Object Interactive Simulation in Real World.' This body presents a four-stage, training-free pipeline: PGSR scene reconstruction, SAM3D object generation, object–scene alignment (physics-constrained pose optimization and mask-guided appearance distillation), and MPM-based multi-object simulation. There is no mention of CA-World, counterfactual states, inverse interventions, or any objective that compares the reintegrated scene to the original under a counterfactual intervention. The experimental section evaluates SAM3D-Phys on six real-world scenes against Feature Splatting and DecoupledGaussian, with a 12-participant user study and LMM-as-judge scores. As submitted, the paper's stated central contribution is entirely absent from the body.

Significance. If the CA-World counterfactual-alignment formulation were actually developed and experimentally validated, it could offer a new way to train interaction-ready reconstruction with direct supervision from counterfactual consistency, potentially improving efficiency by avoiding joint optimization of all object states. However, the manuscript as submitted does not deliver this: the abstract claims a method that the full text does not describe. The body's SAM3D-Phys pipeline is a plausible engineering contribution—combining a generative 3D prior with reconstruction and physics-based constraints—but it is a different contribution, and its own evidence (six scenes, two baselines, no error bars, small user study) is thin. The mismatch between the abstract and body prevents assessment of the claimed novelty and precludes acceptance. The manuscript contains machine-checkable equations only for standard 3DGS, MPM, and basic penalty constraints; the central counterfactual consistency objectives have no formalization.

major comments (3)
  1. [Abstract and Section 3 (whole body)] The abstract claims CA-World counterfactual alignment learning with three objectives between reintegrated and original scenes, and formulates foreground-background decoupling as a visual intervention, object generation/inpainting as counterfactual generation, and reintegration as an inverse intervention. The full text contains none of this. Section 3.2 presents a pipeline (PGSR, SAM3D, alignment, MPM) with no mention of counterfactual states; Sections 3.3–3.4 give render-and-compare pose optimization (Eq. 5), physics penalty constraints (Eqs. 6–8), and a mask-guided VGG distillation loss, none of which are derived from or compared against a counterfactual 'inverse intervention.' No equation, algorithm, or experiment implements the abstract's central claim. The Limitations (Fig. 9) only discuss appearance/context failures, not this missing formalization. This is a load-bearing absence: th
  2. [Abstract (counterfactual consistency formulation)] The abstract's formulation is circular if interpreted literally. It states that 'reversing the intervention should recover the factual world' and motivates 'three alignment objectives between the reintegrated and original scenes.' If the reintegrated scene is directly supervised against the original observed scene, then the 'counterfactual state' is matched to the factual observation it is supposed to counterfactually predict—this is not causal consistency but regression to the factual scene. Since the body does not formalize these objectives, the claim is untestable as stated. The reader is left with either a circular training target or an undefined one; neither supports the claimed novelty.
  3. [Section 4 (Experiments)] Even for the body's more modest SAM3D-Phys contribution, the quantitative evidence is thin. The evaluation uses six scenes (two from DecoupledGaussian, four self-captured), two baselines, and a 12-participant user study with no reported variance, inter-rater reliability, or significance tests. Table 1 reports single ablation numbers for edge error, PSNR, and SSIM with no error bars. Baselines O2-Recon and Amodal3R are excluded with the justification that performance on real scenes is poor—this is not a substitute for a comparative experiment. These limitations undermine the claim of 'demonstrate effectiveness' (Section 4.2) for the body's method.
minor comments (6)
  1. [Eq. (5)] The notation in Eq. (5) is tangled: subscripts and superscripts on t and R are difficult to parse, and the use of '⊙' for rotation composition is nonstandard. Please clarify.
  2. [Abstract vs. body project page] The abstract gives the project page as https://chnxindong.github.io/ca-world/, while the body (title and footer) gives https://chnxindong.github.io/sam3d-phys/. These should match.
  3. [Fig. 8 caption] The caption contains 'massgase gun,' which appears to be a typo for 'massage gun.'
  4. [References] O2-Recon and Amodal3R are cited twice with different reference numbers (13/14 and 40/41 respectively); please use a single reference per work.
  5. [Eq. (7) and Supplementary Equation (7)] Equation (7) uses a minimum distance margin 'm,' while the supplementary text says the tolerance can be defined as zero and uses 'epsilon.' This symbol inconsistency should be resolved.
  6. [Table 2 (User study)] The user study reports only aggregate scores. Please report per-participant variance or inter-rater agreement to support the claim that the method 'significantly outperforms' baselines.

Circularity Check

1 steps flagged

Abstract's counterfactual-consistency objective is defined as matching the reintegrated scene to the original observed scene, making it a reconstruction loss; the full text omits the claimed CA-World method entirely.

specific steps
  1. self definitional [Abstract, 'counterfactual consistency' paragraph]
    "According to counterfactual consistency, reversing the intervention should recover the factual world, motivating three alignment objectives between the reintegrated and original scenes: appearance, spatial, and physical consistency. This formulates interaction-ready reconstruction as counterfactual alignment learning with direct supervision."

    The counterfactual state is defined as the reintegrated scene, and consistency is defined as agreement with the original observed scene. The supervision for the counterfactual is therefore the very factual observation it is supposed to counterfactually predict; 'reversing the intervention should recover the factual world' is an autoencoding/reconstruction loss, not an independent causal check. By construction, any reintegration fitted to the original scene trivially satisfies the stated objective, so the claimed counterfactual alignment reduces to reconstruction.

full rationale

The only concrete circular reduction is in the abstract: the three 'counterfactual consistency' objectives are defined as matching the reintegrated scene to the original observed scene, which makes the counterfactual constraint equivalent to a reconstruction/autoencoding objective by construction. The full text compounds this by omitting the claimed CA-World machinery entirely: the body is a different paper, SAM3D-Phys, whose alignment modules fit object pose and texture to the observed images via render-and-compare and mask-guided VGG distillation. Those are straightforward supervised fitting steps, not counterfactual predictions, and no equation in the body defines the abstract's appearance, spatial, or physical consistency objectives under an inverse intervention. There is no load-bearing self-citation chain; the self-citation to Pos3R is only related work, and the generative prior SAM3D is an external system. Because the central advertised result is both definitionally circular where stated and absent where the method is described, the circularity score is 6 rather than 2. The body's own experiments are self-contained benchmark comparisons, but they evaluate SAM3D-Phys, not the CA-World claim in the abstract.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 1 invented entities

The central CA-World claim is not implemented, so the ledger mostly catalogs assumptions of the body's SAM3D-Phys. The abstract's counterfactual-consistency premise is treated as an ad hoc axiom; the body relies on known 3DGS/MPM math and on unvalidated domain assumptions about SAM3D, LaMa, PGSR depths, and Gemini relation graphs.

free parameters (4)
  • per-object pose (translation t, rotation R) = optimized per scene; values not reported
    Eq. 5 iteratively fits t and R by render-and-compare against ground-truth masked images; these are fitted degrees of freedom, not predictions.
  • physical-constraint tolerance epsilon and margin m = epsilon 'small predefined tolerance'; m unspecified
    Eqs. 6-7 introduce hand-chosen thresholds deciding what counts as grounded and non-interpenetrating placement.
  • MPM material parameters = inherited from PhysGaussian/DreamPhysics; not measured
    Section 3.5 and Suppl. E use assigned mass, deformation gradient and material properties without estimating them from data.
  • background inpainting / PGSR fine-tune settings = not reported
    Section 3.2 fine-tunes the reconstructed background after LaMa inpainting; hyperparameters are unspecified.
axioms (6)
  • standard math 3D Gaussian Splatting forward model (Eq. 1) and pinhole back-projection (Eq. 4) are correct.
    Used throughout Section 3 without proof; standard rendering/geometry results.
  • domain assumption PGSR unbiased depth (Eq. 3) approximates the physical surface.
    Section 3.1; depth feeds pointmaps used for SAM3D initialization and metric alignment; biased depth would break alignment.
  • domain assumption SAM3D infers complete object geometry from partial observations consistently with the scene.
    Section 3.2; core generative-prior assumption, not validated against ground-truth object scans.
  • ad hoc to paper LaMa inpainting plus PGSR fine-tuning yields a clean background suitable as a counterfactual state.
    Section 3.2; background removal/inpainting is treated as the reference for the 'counterfactual' scene.
  • ad hoc to paper Counterfactual consistency: reversing the intervention should recover the factual world.
    Abstract only; no formal definition in body; motivates the three alignment objectives.
  • domain assumption Gemini 3 Pro relation graph correctly identifies object-object and object-scene contacts.
    Section 3.3, Fig. 2; physical constraints rely on this external VLM output.
invented entities (1)
  • counterfactual state no independent evidence
    purpose: target of CA-World's alignment objectives and efficiency claim
    Mentioned only in the abstract; no formal representation, equation, or falsifiable prediction in the full text.

reviewed 2026-08-04 · how reviews work

0 comments
read the original abstract

Reconstructing interaction-ready 3D worlds is essential for physical simulation, virtual reality, robotics, and autonomous driving. However, existing methods mainly optimize static and holistic visual fidelity, with limited support for multi-object interaction. We argue that an interaction-ready reconstruction should anticipate potential scene changes and preserve geometric completeness, visual quality, multi-object spatial relationship, and physical plausibility under potential interactions. To this end, motivated by the causal intervention, we propose CA-World, an efficient framework that integrates counterfactual alignment learning into a decoupling-reintegration reconstruction pipeline. Specifically, we formulate foreground-background decoupling as a visual intervention, separate object generation and background inpainting as counterfactual generation, and scene reintegration as an inverse intervention. According to counterfactual consistency, reversing the intervention should recover the factual world, motivating three alignment objectives between the reintegrated and original scenes: appearance, spatial, and physical consistency. This formulates interaction-ready reconstruction as counterfactual alignment learning with direct supervision. Moreover, leveraging the locality of object-level interventions, CA-World constrains counterfactual states using the observed scene, enabling efficient and coherent reintegration without jointly optimizing all object states, thereby reducing computational cost and error accumulation. Experiments on object completeness, spatial accuracy, outdoor background completion, rendering quality, simulated dynamics, and downstream applications demonstrate the effectiveness of CA-World. Project page: https://chnxindong.github.io/ca-world/.

Figures

Figures reproduced from arXiv: 2605.30239 by Lihan Zhang, Tianru Dai, Weijian Deng, Wenfeng Deng, Xin Dong, Yansong Tang.

Figure 1
Figure 1. Figure 1: SAM3D-Phys Overview. The pipeline consists of four major steps: (A) Scene reconstruction, where the scene is reconstructed from multi-view images using PGSR, followed by object removal and inpainting to obtain a clean background scene; (B) Ob￾ject extraction, where target objects are segmented and converted into complete 3D ge￾ometry using image-to-3D generation with SAM3D; (C) Object–scene alignment, wher… view at source ↗
Figure 2
Figure 2. Figure 2: Physics-constrained alignment refinement. A control graph is constructed to model object–object and scene–object relations, which guide spatial refinement and physically consistent final object placement. 1) object-scene relation constraint, which ensures the object maintains stable contact with the ground without floating or sinking. This can be written as: \mathcal {L}_{\text {os}} = \frac {1}{|\mathcal … view at source ↗
Figure 3
Figure 3. Figure 3: Comparisons with the state of the arts. Our method ensures accurate decou￾pling, alignment, and fidelity. Feature Splatting suffers from fragmentation due to poor background separation. DecoupledGaussian separates foregrounds but fails to disentan￾gle individual objects. 3.5 Interactive Simulation After the restored objects are reintegrated into the scene, their dynamic in￾teractions are simulated using a … view at source ↗
Figure 4
Figure 4. Figure 4: Subfigure (a) shows the results of decoupled object recory. Subfigure (b) illus￾trates the results of physical constrained optimization, where the object is well-aligned with the mask region and maintains a stable position without interpenetration [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: With the pointmaps, generative model can achieve more accurate object ini￾tialization, which helps better spatial alignment and stable physical simulation. 4.2 Comparison with the State of the Art Rendered videos. To validate the multi-object decoupling and appearance consistency for physical simulation, we show comparisons with Feature Splat￾ting and DecoupledGaussian on the room example. As shown in [PI… view at source ↗
Figure 6
Figure 6. Figure 6: Two examples of render-and-compare refinement for spatial alignment. Each column shows the intermediate result at increasing optimization steps. (a) Example of mask-based alignment for the bear statue, where the rendered object progressively aligns with the observed mask region in the image. (b) Example of pose refinement for the ball object, where the object position gradually converges to the correct pla… view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of appearance alignment. The second row shows the object before alignment and the third row after alignment, demonstrating improved texture details and closer visual consistency with the in-scene object [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of the physical simulation and dynamic scene editing. Subfigure(a) shows physical simulation with “massgase gun"(object), while Subfigure(b) shows phys￾ical simulation after removing “massgase gun". 4.4 Evaluating Object-Scene Appearance Alignment To ensure that extracted objects are accurate not only in geometry and spatial placement but also in visual appearance, we introduce a mask-based a… view at source ↗
Figure 9
Figure 9. Figure 9: Limitations caused by appearance variations (e.g., specular highlights and shad￾ows) and less-informative contextual cues. show that without appearance distillation, the PSNR and SSIM metrics degrade significantly, indicating a substantial drop in visual fidelity. Applications. As demonstrated above, our method achieves high-quality multi￾object spatial and appearance alignment. Coupled with Material Point… view at source ↗
Figure 10
Figure 10. Figure 10: Zoomed-in comparisons with baselines. Unlike baseline approaches dependent on natural language for localization and segmentation, which fail in multi-object sce￾narios and suffer from severe tearing and blur, our method utilizes generative priors to resolve object separation gaps in 3D reconstruction. By jointly exploiting metric cues and appearance information from the reconstruction process, our method … view at source ↗
Figure 11
Figure 11. Figure 11: Dataset scenes used in our experiments. These six real-world multi-object scenes exhibit object coupling effects, such as mutual occlusions, making them suited for evaluating the effectiveness of our method in object separation, spatial alignment, and appearance alignment. Outdoor This scene presents an outdoor park setting, where various objects, including a massage gun and a pink speaker, are arranged o… view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Occlusion-Robust Multi-Object Decoupling for Physics-Based Robotic Interaction

    cs.CV 2026-06 unverdicted novelty 4.0

    A pipeline combining SAM2 segmentation, 3D Gaussian Splatting, and joint Score Distillation Sampling with 2D/3D diffusion priors reconstructs decoupled multi-object geometries from occluded sparse views for MPM simulation.

  2. Occlusion-Robust Multi-Object Decoupling for Physics-Based Robotic Interaction

    cs.CV 2026-06 unverdicted novelty 4.0

    A new pipeline for occlusion-robust multi-object 3D reconstruction from sparse views supports physics-based robotic interaction.

Reference graph

Works this paper leans on

48 extracted references · 15 linked inside Pith · cited by 1 Pith paper

  1. [1]

    In: ICML (2024)

    Bruce, J., Dennis, M.D., Edwards, A., Parker-Holder, J., Shi, Y., Hughes, E., Lai, M., Mavalankar, A., Steigerwald, R., Apps, C., et al.: Genie: Generative interactive environments. In: ICML (2024)

    2024

  2. [2]

    Cai, D., Heikkilä, J., Rahtu, E.: Gs-pose: Generalizable segmentation-based 6d object pose estimation with 3d gaussian splatting (2024)

    2024

  3. [3]

    In: CVPR

    Cao, T., Luo, F., Qin, J., Jiang, Y., Wang, Y., Xiao, C.: ig-6dof: Model-free 6dof pose estimation for unseen object via iterative 3d gaussian splatting. In: CVPR. pp. 6436–6446 (2025)

    2025

  4. [4]

    IEEE TVCG31, 6100–6111 (2024)

    Chen, D., Li, H., Ye, W., Wang, Y., Xie, W., Zhai, S., Wang, N., Liu, H., Bao, H., Zhang, G.: Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. IEEE TVCG31, 6100–6111 (2024)

    2024

  5. [5]

    arXiv preprint arXiv:2511.16624 (2025)

    Chen, X., Chu, F.J., Gleize, P., Liang, K.J., Sax, A., Tang, H., Wang, W., Guo, M., Hardin, T., Li, X., et al.: Sam 3d: 3dfy anything in images. arXiv preprint arXiv:2511.16624 (2025)

    Pith/arXiv arXiv 2025

  6. [6]

    In: CVPR

    Deng, W., Campbell, D., Sun, C., Zhang, J., Kanitkar, S., Shaffer, M.E., Gould, S.: Pos3r: 6d pose estimation for unseen objects made easy. In: CVPR. pp. 16818– 16828 (2025)

    2025

  7. [7]

    Feng, J., Li, X., Lin, J., Liu, J., Liu, G., Lou, W., Ma, S., Shi, G., Wang, Q., Wang, J., et al.: Seed3d 1.0: From images to high-fidelity simulation-ready 3d assets (2025)

    2025

  8. [8]

    In: CVPR

    Geng, D., Herrmann, C., Hur, J., Cole, F., Zhang, S., Pfaff, T., Lopez-Guevara, T., Aytar, Y., Rubinstein, M., Sun, C., et al.: Motion prompting: Controlling video generation with motion trajectories. In: CVPR. pp. 1–12 (2025)

    2025

  9. [9]

    arXiv preprint arXiv:2307.04725 (2023)

    Guo, Y., Yang, C., Rao, A., Liang, Z., Wang, Y., Qiao, Y., Agrawala, M., Lin, D., Dai, B.: Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725 (2023)

    Pith/arXiv arXiv 2023

  10. [10]

    arXiv preprint arXiv:2311.04400 (2023)

    Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400 (2023)

    Pith/arXiv arXiv 2023

  11. [11]

    arXiv preprint arXiv:2605.15843 (2026)

    Hu, J., Guo, J., Cen, J., Yang, C., Li, S., Shen, W.: Worldact: Activating monolithic 3d worlds into interactive-ready object-centric scenes. arXiv preprint arXiv:2605.15843 (2026)

    Pith/arXiv arXiv 2026

  12. [12]

    ACM Transactions on Graphics (TOG)37(4), 1–14 (2018)

    Hu, Y., Fang, Y., Ge, Z., Qu, Z., Zhu, Y., Pradhana, A., Jiang, C.: A moving least squares material point method with displacement discontinuity and two-way rigid body coupling. ACM Transactions on Graphics (TOG)37(4), 1–14 (2018)

    2018

  13. [13]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Hu, Y., Ye, S., Zhao, W., Lin, M., He, Y., Wen, Y.H., He, Y., Liu, Y.J.: Oˆ 2-recon: Completing3dreconstructionofoccludedobjectsinthescenewithapre-trained2d diffusion model. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 2285–2293 (2024)

    2024

  14. [14]

    In: AAAI

    Hu, Y., Ye, S., Zhao, W., Lin, M., He, Y., Wen, Y.H., He, Y., Liu, Y.J.: O2-recon: completing 3d reconstruction of occluded objects in the scene with a pre-trained 2d diffusion model. In: AAAI. vol. 38, pp. 2285–2293 (2024)

    2024

  15. [15]

    In: Pro- ceedings of the AAAI Conference on Artificial Intelligence

    Huang, T., Zhang, H., Zeng, Y., Zhang, Z., Li, H., Zuo, W., Lau, R.W.: Dream- physics: Learning physics-based 3d dynamics with video diffusion priors. In: Pro- ceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 3733–3741 (2025)

    2025

  16. [16]

    In: ACM SIGGRAPH

    Jiang, Y., Yu, C., Xie, T., Li, X., Feng, Y., Wang, H., Li, M., Lau, H., Gao, F., Yang, Y., et al.: Vr-gs: A physical dynamics-aware interactive gaussian splatting system in virtual reality. In: ACM SIGGRAPH. pp. 1–1 (2024) 16 X. Dong et al

    2024

  17. [17]

    ACM TOG42(4), 139–1 (2023)

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM TOG42(4), 139–1 (2023)

    2023

  18. [18]

    In: IROS

    Kruzliak, A., Hartvich, J., Patni, S.P., Rustler, L., Behrens, J.K., Abu-Dakka, F.J., Mikolajczyk, K., Kyrki, V., Hoffmann, M.: Interactive learning of physical object properties through robot manipulation and database of object measurements. In: IROS. pp. 7596–7603 (2024)

    2024

  19. [19]

    In: CoRL (2022)

    Labbé, Y., Manuelli, L., Mousavian, A., Tyree, S., Birchfield, S., Tremblay, J., Carpentier, J., Aubry, M., Fox, D., Sivic, J.: Megapose: 6d pose estimation of novel objects via render & compare. In: CoRL (2022)

    2022

  20. [20]

    arXiv preprint arXiv:2509.07920 (2025)

    Li, A., Liu, J., Zhu, Y., Tang, Y.: Scorehoi: Physically plausible reconstruc- tion of human-object interaction via score-guided diffusion. arXiv preprint arXiv:2509.07920 (2025)

    Pith/arXiv arXiv 2025

  21. [21]

    arXiv preprint arXiv:2303.05512 (2023)

    Li, X., Qiao, Y.L., Chen, P.Y., Jatavallabhula, K.M., Lin, M., Jiang, C., Gan, C.: Pac-nerf: Physics augmented continuum neural radiance fields for geometry- agnostic system identification. arXiv preprint arXiv:2303.05512 (2023)

    Pith/arXiv arXiv 2023

  22. [22]

    In: CVPR

    Li, Z., Tucker, R., Snavely, N., Holynski, A.: Generative image dynamics. In: CVPR. pp. 24142–24153 (2024)

    2024

  23. [23]

    ICLR (2025)

    Lin, Y., Lin, C., Xu, J., Mu, Y.: Omniphysgs: 3d constitutive gaussians for general physics-based dynamics generation. ICLR (2025)

    2025

  24. [24]

    arXiv preprint arXiv:2406.04338 (2024)

    Liu,F.,Wang,H.,Yao,S.,Zhang,S.,Zhou,J.,Duan,Y.:Physics3d:Learningphys- ical properties of 3d gaussians via video diffusion. arXiv preprint arXiv:2406.04338 (2024)

    Pith/arXiv arXiv 2024

  25. [25]

    In: ECCV

    Liu, S., Ren, Z., Gupta, S., Wang, S.: Physgen: Rigid-body physics-grounded image-to-video generation. In: ECCV. pp. 360–378 (2024)

    2024

  26. [26]

    In: CVPR

    Liu, Z., Ye, W., Luximon, Y., Wan, P., Zhang, D.: Unleashing the potential of multi-modal foundation models and video diffusion for 4d dynamic physical scene simulation. In: CVPR. pp. 11016–11025 (2025)

    2025

  27. [27]

    In: ICRA

    Lou, H., Liu, Y., Pan, Y., Geng, Y., Chen, J., Ma, W., Li, C., Wang, L., Feng, H., Shi, L., et al.: Robo-gs: A physics consistent spatial-temporal model for robotic arm with hybrid representation. In: ICRA. pp. 15379–15386 (2025)

    2025

  28. [28]

    In: ECCV

    Lu, G., Zhang, S., Wang, Z., Liu, C., Lu, J., Tang, Y.: Manigaussian: Dynamic gaussian splatting for multi-task robotic manipulation. In: ECCV. pp. 349–366 (2024)

    2024

  29. [29]

    In: IEEE Conference on Virtual Reality and 3D User Interfaces Abstracts and Workshops (VRW)

    Mao, H., Xu, Z., Wei, S., Quan, Y., Deng, N., Yang, X.: Live-gs: Llm powers interactive vr by enhancing gaussian splatting. In: IEEE Conference on Virtual Reality and 3D User Interfaces Abstracts and Workshops (VRW). pp. 1234–1235 (2025)

    2025

  30. [30]

    CACM 65(1), 99–106 (2021)

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. CACM 65(1), 99–106 (2021)

    2021

  31. [31]

    arXiv preprint arXiv:2404.01223 (2024)

    Qiu, R.Z., Yang, G., Zeng, W., Wang, X.: Feature splatting: Language-driven physics-based scene synthesis and editing. arXiv preprint arXiv:2404.01223 (2024)

    Pith/arXiv arXiv 2024

  32. [32]

    arXiv preprint arXiv:2408.00714 (2024)

    Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C., Gustafson, L., et al.: Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714 (2024)

    Pith/arXiv arXiv 2024

  33. [33]

    arXiv preprint arXiv:1409.1556 (2014)

    Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)

    Pith/arXiv arXiv 2014

  34. [34]

    arXiv preprint arXiv:2109.07161 (2021) SAM3D-Phys: Towards Multi-Object Interactive Simulation in Real World 17

    Suvorov, R., Logacheva, E., Mashikhin, A., Remizova, A., Ashukha, A., Silvestrov, A., Kong, N., Goka, H., Park, K., Lempitsky, V.: Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161 (2021) SAM3D-Phys: Towards Multi-Object Interactive Simulation in Real World 17

    Pith/arXiv arXiv 2021

  35. [35]

    arXiv preprint arXiv:2402.05054 (2024)

    Tang, J., Chen, Z., Chen, X., Wang, T., Zeng, G., Liu, Z.: Lgm: Large multi- view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054 (2024)

    Pith/arXiv arXiv 2024

  36. [36]

    Team, T.H.: Hunyuan3d 2.0: Scaling diffusion models for high resolution textured 3d assets generation (2025)

    2025

  37. [37]

    arXiv preprint arXiv:2509.20358 (2025)

    Wang, C., Chen, C., Huang, Y., Dou, Z., Liu, Y., Gu, J., Liu, L.: Physctrl: Genera- tive physics for controllable and physics-grounded video generation. arXiv preprint arXiv:2509.20358 (2025)

    arXiv 2025

  38. [38]

    In: CVPR

    Wang, M., Zhang, Y., Xu, W., Ma, R., Zou, C., Morris, D.: Decoupledgaussian: Object-scene decoupling for physics-based interaction. In: CVPR. pp. 11361–11372 (2025)

    2025

  39. [40]

    Wu, T., Zheng, C., Guan, F., Vedaldi, A., Cham, T.J.: Amodal3r: Amodal 3d reconstruction from occluded 2d images (2025)

    2025

  40. [41]

    arXiv preprint arXiv:2503.13439 (2025)

    Wu, T., Zheng, C., Guan, F., Vedaldi, A., Cham, T.J.: Amodal3r: Amodal 3d reconstruction from occluded 2d images. arXiv preprint arXiv:2503.13439 (2025)

    Pith/arXiv arXiv 2025

  41. [42]

    NeurIPS38, 32501–32524 (2026)

    Xia, H., Lin, C.H., Hsu, H.Y., Leboutet, Q., Gao, K., Paulitsch, M., Ummenhofer, B.,Wang,S.:Holoscene:Simulation-readyinteractive3dworldsfromasinglevideo. NeurIPS38, 32501–32524 (2026)

    2026

  42. [43]

    arXiv preprint arXiv:2412.01506 (2024)

    Xiang, J., Lv, Z., Xu, S., Deng, Y., Wang, R., Zhang, B., Chen, D., Tong, X., Yang, J.: Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506 (2024)

    Pith/arXiv arXiv 2024

  43. [44]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Xiang, J., Lv, Z., Xu, S., Deng, Y., Wang, R., Zhang, B., Chen, D., Tong, X., Yang, J.: Structured 3d latents for scalable and versatile 3d generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 21469–21480 (2025)

    2025

  44. [45]

    In: CVPR

    Xie, T., Zong, Z., Qiu, Y., et al.: Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In: CVPR. pp. 4389–4398 (2024)

    2024

  45. [46]

    In: Proceedings of the CVPR

    Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Lavt: Language- aware vision transformer for referring image segmentation. In: Proceedings of the CVPR. pp. 18155–18165 (2022)

    2022

  46. [47]

    In: CVPR

    Yu, H.X., Duan, H., Herrmann, C., Freeman, W.T., Wu, J.: Wonderworld: Inter- active 3d scene generation from a single image. In: CVPR. pp. 5916–5926 (2025)

    2025

  47. [48]

    In: ECCV

    Zhang, T., Yu, H.X., Wu, R., et al.: Physdreamer: Physics-based interaction with 3d objects via video generation. In: ECCV. pp. 388–406 (2024)

    2024

  48. [49]

    Dong et al

    Zhang, X., Chen, Y., Fang, Y., Qu, W., Huang, H., Zhang, C., Xu, F., Li, X.: Telephysics: Physics-grounded multi-object scene generation from a single image with real-time interaction (2026),https://arxiv.org/abs/2605.20290 18 X. Dong et al. Supplementary Material A Zoomed-in Comparisons with Baselines To better highlight the differences between our metho...

    Pith/arXiv arXiv 2026

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.