Pith. sign in

REVIEW 5 major objections 6 minor 46 references

UniPhysGen claims that a single model can take arbitrarily decomposed 3D meshes and jointly predict articulation kinematics and intrinsic physical properties without canonicalized structures, and that this yields state-of-the-art results on

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

2026-08-02 04:43 UTC pith:BBVV3TXD

load-bearing objection Useful pipeline and dataset paper, but the headline SOTA claim rests on a benchmark the same pipeline labeled; worth refereeing, not desk-rejecting. the 5 major comments →

arxiv 2607.13586 v2 pith:BBVV3TXD submitted 2026-07-15 cs.CV

UniPhysGen: Unified Physical Grounding for Simulation-Ready 3D Assets

Xian Li , Rong Wei , Lujie Yang , Haolin Huang , Junyuan Fang , Siliang Tang , Jun Xiao , Rui Tang
show 1 more author
Juncheng Li
This is my paper
classification cs.CV MSC 68T4568T07
keywords physical groundingsimulation-ready 3D assetsarticulation semanticsintrinsic physical propertiesheterogeneous part decompositionsgeometry-robust articulation groundingSO(3) augmentationspherical axis parameterization
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.

Most 3D assets are visually realistic but lack the physical semantics needed for simulation, and existing methods either treat articulation and physical properties separately or depend on clean, canonicalized part decompositions. This paper claims a single model, UniPhysGen, can jointly ground articulation and intrinsic physical properties directly on heterogeneous, inconsistently decomposed meshes. It backs this with a four-stage automated pipeline that builds a 40K-asset dataset and a manually verified benchmark, and reports state-of-the-art results across most evaluation settings. The key move is geometric: replacing orientation shortcuts with full rotation augmentation and spherical axis parameterization, plus a globally shared coordinate encoding, which the authors argue makes articulation reasoning robust to arbitrary orientations and messy part abstractions.

Core claim

UniPhysGen formalizes unified physical grounding as a single autoregressive model that maps object and part point clouds to four structured outputs: part-level physical properties, articulation kinematics (joint type, axis, pivot, limits), motion-coupled part groups, and object-level properties (category, dimensions, mass). Its central claim is that this can be done reliably on heterogeneous, non-canonicalized decompositions by (1) using SO(3)-based rotation augmentation instead of z-axis augmentation to prevent geometry-shortcut bias, (2) parameterizing the articulation axis in spherical coordinates to avoid sign-token instability in language-model decoding, and (3) sharing one global voxel

What carries the argument

The load-bearing device is the geometry-robust articulation module: SO(3) rotation augmentation stops the model from latching onto upright-orientation shortcuts, spherical (θ, φ) parameterization of the normalized axis prevents token-level sign flips from destabilizing axis prediction, and a globally shared voxel grid with Fourier positional encoding anchors part-level and object-level geometry in the same coordinate frame (cutting pivot error from 0.319 to 0.099 in ablation). Around this, a physical semantic alignment pretraining stage ties local geometry to structured physical semantic descriptions, and the four-stage UniPhys pipeline—perceptually guided part decomposition, multimodal phys

Load-bearing premise

The benchmark's ground-truth labels are generated by the same pipeline being evaluated and then human-corrected; if the human pass is shallow or inherits pipeline priors—like the prompt that tells annotators to treat an incomplete drawer front as a complete sliding drawer—then reported scores partly measure label propagation rather than independent physical correctness.

What would settle it

Take a random subset of UniPhys-Bench not shown to UniPhysGen, have independent human annotators re-annotate articulation and physical properties from scratch using geometry and material evidence only, then compare predictions against those fresh labels: if accuracy drops substantially, or if the fresh labels disagree with the benchmark's corrected labels beyond a small threshold, the benchmark has encoded the pipeline's priors.

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

If this is right

  • If UniPhysGen generalizes as reported, arbitrarily decomposed 3D assets from artists, CAD, and generative models can be converted into simulation-ready artifacts without manual cleanup or canonicalization.
  • Joint reasoning over articulation and physical properties should yield more physically consistent motions than independent prediction, since mass, friction, and joint limits are anchored to the same geometry.
  • SO(3) augmentation with spherical axis prediction offers a recipe for orientation-robust kinematic reasoning that can transfer to other 3D interaction tasks.
  • Simulation-driven consistency verification acts as a scalable filter: it rejects or refines implausible annotations without needing exact physics, making large-scale dataset construction tractable.
  • The manually verified benchmark, with heterogeneous part decompositions and a challenging incomplete-geometry subset, gives the field a common testbed for unified physical grounding.

Where Pith is reading between the lines

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

  • Because the benchmark labels are initially produced by the same pipeline and only later corrected by humans, the model's scores may partly reflect reinforcement of the pipeline's priors; a fully independent ground-truth pass would likely shrink the reported margins.
  • The general design—geometric candidate generation followed by semantic selection—is a reusable pattern: constrain the continuous space with geometry, then let a language model rank physically feasible options, which could apply to grasp planning, scene editing, and interactive simulation.
  • A cheap falsification test: measure inter-annotator agreement between the benchmark's corrected labels and fresh human annotation of a random subset; if mismatch is high, state-of-the-art comparisons against this benchmark are unreliable.
  • The introduced robustness tests (challenging subset, merged part granularity, arbitrary rotations) establish a template for evaluating any articulation-grounding method, not just this one.

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

5 major / 6 minor

Summary. The paper introduces UniPhys, an automated pipeline that converts raw 3D meshes into simulation-ready assets by decomposing objects, grounding articulation kinematics and intrinsic physical properties, and verifying results through physics simulation. The pipeline is used to build UniPhys-40K, a large training dataset, and UniPhys-Bench, a 1,927-object benchmark with manual verification. The authors also propose UniPhysGen, a unified Qwen3/Sonata-based model that jointly predicts part-level physical properties, articulation kinematics, articulation structure, and object-level properties. The main claim is state-of-the-art performance across articulation and intrinsic-property grounding tasks, with particular robustness to heterogeneous part decompositions. Supplementary material contains detailed pipeline, prompts, metrics, and additional PartNet-Mobility experiments.

Significance. If the claims are sustained, the paper makes a useful engineering contribution: it is among the first to combine articulation semantics and physical properties in a single scalable pipeline for non-canonicalized 3D assets, and the proposed geometry-robust articulation grounding (SO(3) augmentation plus spherical axis parameterization) is a sensible response to a real rotation-bias problem. The dataset, if made public and independently validated, could be a useful resource. The manuscript also includes honest ablations and a clear description of the verification protocol. However, the central SOTA claim rests on a benchmark whose labels are produced by the same pipeline being evaluated, and the external PartNet-Mobility results do not substantiate a general SOTA claim. The significance is therefore conditional on breaking the circularity and adding independent validation.

major comments (5)
  1. [Sec 3.2 / App A.5 / App G.1] The main claim of SOTA performance on UniPhys-Bench is compromised by benchmark circularity. Appendix A.5 states that UniPhys-Bench annotations were 'initially annotated by the UniPhys pipeline and subsequently inspected and corrected by human annotators.' The model UniPhysGen is trained on UniPhys-40K, which is produced by the same pipeline. The benchmark therefore measures reproduction of the pipeline's annotation policy rather than external physical correctness. Furthermore, the LVLM prompt in App G.1 (Listing 9) explicitly instructs annotators to 'Treat visually incomplete parts as complete functional objects' and provides the drawer-front example. This is the same convention celebrated in Sec 5.3. No inter-annotator agreement, correction-depth statistics, or evidence that human annotators overrode pipeline priors is reported. As a result, Tables 2, 3, 4, and the associated SOTA clai
  2. [Sec 5.1 / Tables 2, 3, 6, 7] No error bars or statistical significance tests are reported anywhere. Many comparisons are close (e.g., Table 2: UniPhysGen vs Articulate-Anything on joint type, 89.96 vs 94.09; Table 6: image-modality ablation differences of 0.01-0.03 in ALDE/MAE). Without variance estimates, multiple seeds, or significance testing, the 'consistently outperforms' statement in Sec 5.2 is not supported. Please provide at least three seeds with mean/std for the main tables, or state why the reported single-run numbers are deterministic.
  3. [Sec 5.2 / Table 2 and App F.1 / Table 9] The intrinsic-physical-property comparison uses only NeRF2Physics as a baseline. The paper does not compare with other physical-property estimators or physics-grounded generation frameworks such as PhysX-3D or PhysX-Anything, which are cited in the related work. On the external PartNet-Mobility benchmark (Table 9), UniPhysGen is not SOTA: it trails PARTICULATE on joint accuracy (98.80 vs 100.00), axis error (8.84 vs 0.52), and pivot error (0.084 vs 0.023), and only leads on limit mIoU. The discussion in App F.1 describes this as 'comparable,' but the main-text claim of SOTA across 'most evaluation settings' is not supported by the external evaluation. Please either qualify the SOTA claim to the specific UniPhys-Bench setting, or provide additional external validation where UniPhysGen is SOTA.
  4. [App G.1 / Sec 5.3] The 'Challenging' subset used to demonstrate robustness is constructed using the same convention that incomplete drawer fronts should be labeled as prismatic (Sec 5.3: 'drawer front panels without underlying drawer bodies'). The evaluation therefore rewards the model for matching the annotation convention injected into the pipeline. Because the benchmark labels inherit this convention (see App A.5), the improvement from 50.09% to 80.53% on Challenging-Rotated under SO(3) augmentation may partly reflect learning the annotation policy rather than learning physically correct articulation. A clean test would be to evaluate on real scanned or independently annotated assets where incomplete geometry is actually broken or ambiguous.
  5. [Sec 3.2 / Fig. 2] The human evaluation in Fig. 2 compares error rates of pipeline outputs before and after simulation-driven verification. This measures whether verification filters or corrects the pipeline's own outputs; it does not establish that the final benchmark labels are externally correct. The claim 'supports the effectiveness of simulation-driven consistency verification' is too strong because the annotators see the pipeline's labels and may simply agree with the verification outcome. Please clarify that this is an internal quality-control measure, not external validation.
minor comments (6)
  1. [Table 1] The Scalability columns S1-S3 are defined in the caption, but the table body for PhysXNet and PhysDB has duplicate entries (e.g., '✗ ✗ ✗' and '✗ ✗' in the same row). Please clean the table formatting so each row has exactly five cells after the first three columns.
  2. [Sec 5.2 / Table 2 caption] The note for Table 2 says 'Note: Scale and affordance are reported for GT and Ours only; baseline results are unavailable.' This appears to be an artifact from a previous version or a misplaced note, since Fig. 7 also mentions baseline unavailability. Please clarify which entries have no baseline and why.
  3. [Sec 4.2 / Eq. (2)] Equation (2) uses notation 'j pi / s k' and '−g obj min' which is not defined until App B.2.4. Please define the bracket notation (flooring) and the voxel grid size s in the main text to avoid confusion.
  4. [App D] The metric definitions for axis angular error use the absolute inner product, which is fine, but the pivot-distance metric is written with a double norm that is easy to misread. Please simplify notation, e.g., use \|\cdot\|_2 explicitly.
  5. [App G.1 / Listing 9] The prompt includes the instruction 'Do not assign neighbors or parent-child links to visually similar but spatially different parts.' This is reasonable, but it also demonstrates that the prompt heavily constrains the LVLM. Please disclose the prompt version used for the final dataset and whether the same prompt was used for human verification.
  6. [Sec 4.2 / Table 3] Table 3 shows that SO(3)-based rotation + spherical augmentation improves axis error under Challenging-Rotated compared to SO(3)-based rotation alone, but still degrades from 6.28 to 23.40 angular error under Rotated-Z. The main text says 'more robust axis estimation' but the absolute errors are still high. Please discuss this residual degradation explicitly.

Circularity Check

2 steps flagged

UniPhys-Bench is seeded by the evaluated UniPhys pipeline, so UniPhysGen's SOTA scores partially measure reproduction of UniPhys annotation conventions rather than independent physical grounding.

specific steps
  1. fitted input called prediction [Section 3.2 (Dataset Overview); Appendix A.5 (UniPhys-Bench Construction)]
    "For both subsets, articulation parameters and intrinsic physical properties are initially annotated by the UniPhys pipeline and subsequently inspected and corrected by human annotators."

    The benchmark used for the headline SOTA claim is an output of the very pipeline being evaluated: UniPhys-Bench labels are 'initially annotated by the UniPhys pipeline,' and UniPhysGen is trained on UniPhys-40K 'driven by the proposed UniPhys pipeline.' High test scores therefore measure how well the model reproduces UniPhys's own annotation policy, not external physical correctness. The reported human 'inspection and correction' does not break the loop without evidence of independent, substantive correction; no inter-annotator agreement, correction-depth audit, or demonstration that annotators overrode pipeline priors is provided. The independent PartNet-Mobility evaluation (Table 9) actually shows UniPhysGen below PARTICULATE on joint accuracy (98.80 vs 100.00), axis error (8.84 vs 0.52)

  2. self definitional [Appendix G.1, Listing 9 (Intrinsic Physical Property Grounding Prompt, Special Rules)]
    "Treat visually incomplete parts as complete functional objects when estimating their motion type or physical parameters. For example, even if only the front panel of a drawer is visible, it should still be considered a complete drawer with its expected sliding motion."

    The ground-truth convention celebrated in Sec. 5.3—drawer front panels without bodies are prismatic—is inserted into the annotation prompt before labels are produced. The LVLM is instructed to treat the incomplete drawer front as a complete drawer with sliding motion, and human annotators inspecting those labels are given the same instruction. The subsequent 'Challenging' evaluation (Sec. 5.3) then reports UniPhysGen correctly infers prismatic motion for such panels. This is a self-definitional loop: the label is defined by the prompt convention, the model is trained to reproduce that convention, and the evaluation confirms agreement with it. It does not establish that the prismatic label is physically correct, and it cannot be used as independent evidence of the method's robustness.

full rationale

The central circularity is in benchmark construction rather than in the model equations. UniPhys-40K is produced by the UniPhys pipeline, UniPhysGen is trained on that dataset, and UniPhys-Bench—the primary evaluation set—is also initially annotated by the same UniPhys pipeline before human correction. Because the paper reports no independent audit of the human correction (no inter-annotator agreement, no quantification of corrections, no evidence that annotators rejected pipeline priors), the SOTA results on UniPhys-Bench partly reflect agreement between the trained model and the label-generating pipeline. This is compounded by the explicit prompt instruction in App. G.1 that defines the drawer-front-as-prismatic convention, which is exactly the behavior later celebrated in Sec. 5.3. The independent PartNet-Mobility evaluation does not rescue the SOTA claim: on the external benchmark, UniPhysGen trails PARTICULATE on joint accuracy, axis error, and pivot error. However, the paper does contain some genuinely independent elements: human verification is claimed (even if not audited), the designer-created Manycore subset adds external structure, and the Isaac Sim deployments are external demonstrations rather than fitted evaluation. No load-bearing self-citation chain was found; the related-work citations are external to the authors' own derivation. Overall, the headline claim is substantially supported only by a benchmark whose ground truth is pipeline-seeded, so the circularity score is 6 rather than a lower score.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

No new physical entities, forces, or mediators are postulated. The load-bearing assumptions are about the reliability of the annotation pipeline: SAM-based perceptual priors, hand-set simulation thresholds, human-corrected benchmark labels, and LVLM-inferred physical properties. Several hand-tuned thresholds directly control which annotations enter UniPhys-40K and UniPhys-Bench, so they are free parameters of the central empirical claim.

free parameters (5)
  • Mass consistency epsilon threshold = 2
    App A.4.1 Level 4: annotations with relative mass inconsistency epsilon>2 are filtered or corrected; threshold chosen to balance coverage and plausibility, directly influences dataset labels.
  • Penetration stability threshold gamma = 3
    App A.4.2 Level 2: Percentile95(contact thickness ratio)>3 invalidates an articulation hypothesis; hand-set threshold used for dataset filtering.
  • Contact identity overlap threshold eta = 0.4
    App A.4.2 Level 3: Percentile10(contact overlap)<0.4 invalidates an articulation; hand-set threshold influences which annotations survive into training/benchmark.
  • Global volume scaling factor = unspecified
    App A.4.1: 'A global scaling factor is then applied to align mesh-space volume with real-world object dimensions'; value is not reported, and it affects the mass-consistency check.
  • Material prior intervals and empirical ranges = material-dependent ranges
    App A.4.1 Levels 1-3: plausible intervals for density/friction/Young's modulus, hardness ratio [1e-3,1e-1], and friction threshold mu>=0.2 are hand-set and filter annotations.
axioms (6)
  • domain assumption Multi-view SAM perceptual grouping coincides with physically meaningful part structure.
    Sec 3.1: 'perceptual grouping priors learned by segmentation foundation models often coincide with physically meaningful structures'; if false, the entire decomposition stage is misdirected.
  • domain assumption MuJoCo-based simulation checks (contact preservation, no penetration, contact identity) are sufficient to certify physical correctness of articulation.
    App A.4.2: verification uses hand-set thresholds and idealizes simulation; annotations that pass are treated as physically consistent for dataset construction.
  • domain assumption Human annotator correction yields reliable ground truth.
    Sec 3.2/App A.5: initial pipeline annotations are 'inspected and corrected by human annotators'; no inter-annotator agreement or correction-depth statistics are reported.
  • domain assumption Intrinsic physical properties are inferable from rendered geometry and appearance plus contextual prompts.
    UniPhys uses GPT-5 to infer density, friction, mass, and material from part-centric renderings; the quality of the dataset depends on this visual reasoning assumption.
  • ad hoc to paper Voxelized mesh volume with a global scaling factor approximates true part volume.
    App A.4.1: volume is estimated by voxelizing each part at 128^3 and applying an unspecified global scaling factor; no validation against known volumes is given.
  • ad hoc to paper Treating visually incomplete parts as complete functional objects is valid for ground-truth annotation.
    App G.1 Special Rules 1-3: annotators and model are instructed to infer motion from functional semantics rather than visible geometry, injecting the paper's desired priors into labels.

pith-pipeline@v1.3.0-alltime-deepseek · 31963 in / 13073 out tokens · 130994 ms · 2026-08-02T04:43:32.658982+00:00 · methodology

0 comments
read the original abstract

Physically grounded 3D assets are increasingly important for embodied AI and robotic simulation. However, most existing 3D assets lack unified physical semantics, including articulation semantics and intrinsic physical properties, required for realistic interaction. Current approaches either treat these semantics independently or rely on canonicalized object structures, limiting robustness across heterogeneous 3D assets. We present UniPhys, a scalable framework for automatically transforming raw 3D assets into simulation-ready assets with unified physical semantics. Based on UniPhys, we construct UniPhys-40K, a large-scale physically grounded dataset, together with UniPhys-Bench, a carefully verified benchmark for unified physical grounding evaluation. We further introduce UniPhysGen, a unified physical grounding model that jointly reasons over articulation semantics and intrinsic physical properties. UniPhysGen incorporates geometry-robust articulation grounding to mitigate geometric shortcut bias under heterogeneous part decompositions. Extensive experiments demonstrate state-of-the-art performance across articulation grounding and intrinsic physical property estimation tasks, while the resulting assets can be directly deployed in robotic simulation environments for realistic physical interaction. Our code and dataset will be available at https://github.com/breezexian/UniPhysGen.

Figures

Figures reproduced from arXiv: 2607.13586 by Haolin Huang, Juncheng Li, Jun Xiao, Junyuan Fang, Lujie Yang, Rong Wei, Rui Tang, Siliang Tang, Xian Li.

Figure 1
Figure 1. Figure 1: UniPhys: A scalable pipeline for transforming heterogeneous 3D assets into simulation [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Human evaluation. a large-scale dataset with grounded articulation se￾mantics and intrinsic physical properties. UniPhys￾40K is built from diverse 3D repositories, including Objaverse Deitke et al. (2023b), HSSD Khanna et al. (2024), 3D-FUTURE Fu et al. (2021), ABO Collins et al. (2022), and PartNet Mo et al. (2019). To support rigorous evaluation, we further construct UniPhys-Bench, a carefully verified b… view at source ↗
Figure 3
Figure 3. Figure 3: Statistics of UniPhys-40K and UniPhys-Bench. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of UniPhysGen. The model takes 3D object and part geometry as input (Left) [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Axis direction reasoning space. To mitigate this issue, we replace canonical￾orientation augmentation with SO(3)-based rotation augmentation ( [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison of articulation grounding on UniPhys-Bench. “Reset” denotes the [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison of intrinsic physical property grounding on UniPhys-Bench. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative simulation-based validation of UniPhysGen-grounded assets. Each sequence [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Details of Perceptually guided physically meaningful structural decomposition pipeline. [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Visualization of context-preserving part-centric renderings for intrinsic physical property [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Visualization of the LVLM input for geometry-aware articulation grounding. For each [PITH_FULL_IMAGE:figures/full_fig_p016_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Additional qualitative comparisons of articulation grounding on UniPhys-Bench. “Reset” [PITH_FULL_IMAGE:figures/full_fig_p031_12.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

46 extracted references · 9 linked inside Pith

  1. [1]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Gaussianproperty: Integrating physical properties to 3d gaussians with lmms , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  2. [2]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Physical property understanding from language-embedded feature fields , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  3. [3]

    International Conference on Learning Representations , volume=

    Real2code: Reconstruct articulated objects via code generation , author=. International Conference on Learning Representations , volume=

  4. [4]

    arXiv preprint arXiv:2502.02590 , year=

    Articulate anymesh: Open-vocabulary 3d articulated objects modeling , author=. arXiv preprint arXiv:2502.02590 , year=

    Pith/arXiv arXiv

  5. [5]

    arXiv preprint arXiv:2512.11798 , year=

    Particulate: Feed-Forward 3D Object Articulation , author=. arXiv preprint arXiv:2512.11798 , year=

    arXiv

  6. [6]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Particulate: Feed-forward 3d object articulation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  7. [7]

    International Conference on Learning Representations , volume=

    Articulate-anything: Automatic modeling of articulated objects via a vision-language foundation model , author=. International Conference on Learning Representations , volume=

  8. [8]

    arXiv preprint arXiv:2405.11656 , year=

    Urdformer: A pipeline for constructing articulated simulation environments from real-world images , author=. arXiv preprint arXiv:2405.11656 , year=

    Pith/arXiv arXiv

  9. [9]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Structured 3d latents for scalable and versatile 3d generation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  10. [10]

    arXiv preprint arXiv:2512.14692 , year=

    Native and compact structured latents for 3d generation , author=. arXiv preprint arXiv:2512.14692 , year=

    Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:2501.12202 , year=

    Hunyuan3d 2.0: Scaling diffusion models for high resolution textured 3d assets generation , author=. arXiv preprint arXiv:2501.12202 , year=

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:2506.16504 , year=

    Hunyuan3d 2.5: Towards high-fidelity 3d assets generation with ultimate details , author=. arXiv preprint arXiv:2506.16504 , year=

    Pith/arXiv arXiv

  13. [13]

    Seed3D 2.0: Advancing High-Fidelity Simulation-Ready 3D Content Generation , author=

  14. [14]

    Advances in Neural Information Processing Systems , volume=

    Objaverse-xl: A universe of 10m+ 3d objects , author=. Advances in Neural Information Processing Systems , volume=

  15. [15]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Objaverse: A universe of annotated 3d objects , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  16. [16]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Abo: Dataset and benchmarks for real-world 3d object understanding , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  17. [17]

    International Journal of Computer Vision , volume=

    3d-future: 3d furniture shape with texture , author=. International Journal of Computer Vision , volume=. 2021 , publisher=

    2021

  18. [18]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  19. [19]

    arXiv preprint arXiv:1512.03012 , year=

    Shapenet: An information-rich 3d model repository , author=. arXiv preprint arXiv:1512.03012 , year=

    Pith/arXiv arXiv

  20. [20]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  21. [21]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Sapien: A simulated part-based interactive environment , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  22. [22]

    arXiv preprint arXiv:2605.05163 , year=

    PhysForge: Generating Physics-Grounded 3D Assets for Interactive Virtual World , author=. arXiv preprint arXiv:2605.05163 , year=

    Pith/arXiv arXiv

  23. [23]

    Advances in Neural Information Processing Systems , volume=

    Physx-3d: Physical-grounded 3d asset generation , author=. Advances in Neural Information Processing Systems , volume=

  24. [24]

    PhysX-3D: Physical-Grounded 3D Asset Generation , volume =

    Cao, Ziang and Chen, Zhaoxi and Pan, Liang and Liu, Ziwei , booktitle =. PhysX-3D: Physical-Grounded 3D Asset Generation , volume =

  25. [25]

    Advances in Neural Information Processing Systems , pages =

    Cao, Ziang and Chen, Zhaoxi and Pan, Liang and Liu, Ziwei , title =. Advances in Neural Information Processing Systems , pages =

  26. [26]

    arXiv preprint arXiv:2511.13648 , year=

    PhysX-Anything: Simulation-Ready Physical 3D Assets from Single Image , author=. arXiv preprint arXiv:2511.13648 , year=

    arXiv

  27. [27]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

    S2o: Static to openable enhancement for articulated 3d objects , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

  28. [28]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Gapartnet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  29. [29]

    International Conference on Learning Representations , volume=

    Singapo: Single image controlled generation of articulated parts in objects , author=. International Conference on Learning Representations , volume=

  30. [30]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Partfield: Learning 3d feature fields for part segmentation and beyond , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  31. [31]

    arXiv preprint arXiv:2509.06784 , year=

    P3-sam: Native 3d part segmentation , author=. arXiv preprint arXiv:2509.06784 , year=

    arXiv

  32. [32]

    International Conference on Learning Representations , volume=

    Sam 2: Segment anything in images and videos , author=. International Conference on Learning Representations , volume=

  33. [33]

    URDF-Anything: Constructing Articulated Objects with 3D Multimodal Language Model , volume =

    Li, Zhe and Bai, Xiang and Zhang, Jieyu and Wu, Zhuangzhe and Xu, Che and Li, Ying and Hou, Chengkai and Zhang, Shanghang , booktitle =. URDF-Anything: Constructing Articulated Objects with 3D Multimodal Language Model , volume =

  34. [34]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    PhysSplat: Efficient Physics Simulation for 3D Scenes via MLLM-Guided Gaussian Splatting , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  35. [35]

    Fast Algorithms for Mining Association Rules in Large Databases , booktitle =

    Rakesh Agrawal and Ramakrishnan Srikant , editor =. Fast Algorithms for Mining Association Rules in Large Databases , booktitle =. 1994 , timestamp =

    1994

  36. [36]

    Naval research logistics quarterly , volume=

    The Hungarian method for the assignment problem , author=. Naval research logistics quarterly , volume=. 1955 , publisher=

    1955

  37. [37]

    arXiv preprint arXiv:2509.08643 , year=

    X-part: high fidelity and structure coherent shape decomposition , author=. arXiv preprint arXiv:2509.08643 , year=

    arXiv

  38. [38]

    2026 , howpublished =

    2026

  39. [39]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    Pith/arXiv arXiv

  40. [40]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Sonata: Self-supervised learning of reliable point representations , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  41. [41]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  42. [42]

    arXiv preprint arXiv:2601.03267 , year=

    Openai gpt-5 system card , author=. arXiv preprint arXiv:2601.03267 , year=

    Pith/arXiv arXiv

  43. [43]

    Advances in Neural Information Processing Systems , volume=

    Spatiallm: Training large language models for structured indoor modeling , author=. Advances in Neural Information Processing Systems , volume=

  44. [44]

    Scaling Learning Algorithms Towards

    Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards

  45. [45]

    and Osindero, Simon and Teh, Yee Whye , journal =

    Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume =

  46. [46]

    2016 , publisher=

    Deep learning , author=. 2016 , publisher=

    2016

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.