Grounding Creativity in Physics: A Brief Survey of Physical Priors in AIGC

Ping Liu; Siwei Meng; Yawei Luo

REVIEW 4 major objections 5 minor 1 cited by

Grounding Creativity in Physics: A Brief Survey of Physical Priors in AIGC

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read This survey establishes a taxonomy of physics-grounded 3D/4D generation and a benchmark showing that Gaussian-splatting methods currently lead on physical-parameter accuracy.

desk verdict Useful taxonomy of physics-aware AIGC, but the one quantitative benchmark table is under-sourced and needs either verification or deletion. read the letter →

arxiv 2502.07007 v3 pith:OVMDWDQ6 submitted 2025-02-10 cs.CV

Siwei Meng , Yawei Luo , Ping Liu This is my paper · ORCID

classification cs.CV

keywords physicspriors3Dgeneration4DNeuralRadianceFieldsGaussianSplattingMaterialPointMethoddifferentiablesimulationAIGC

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Physics priors are becoming a distinct axis in AI-generated 3D and 4D content, and this survey tries to organize that axis with a taxonomy: static 3D generation, dynamic 3D generation, and 4D generation, each subdivided by representation into vision-based, NeRF-based, and Gaussian-splatting-based methods. The motivating problem is that appearance-focused generative models produce artifacts such as unstable deformations, implausible interactions, and physically impossible motion, because they never represent mass, stiffness, or forces. The survey argues that differentiable physics engines, chiefly the Material Point Method, the Finite Element Method, and the DiffTaichi framework, can be inserted into these pipelines to enforce structural integrity and motion realism. Its comparative benchmark on the Synthetic PAC-NeRF dataset finds that no single method wins across all materials: one Gaussian-splatting method is most robust for fluids and plasticine, another leads on elastic objects, and NeRF-based methods trail because of first-frame initialization. A sympathetic reader should take away that representation choice and solver choice are now empirical, material-dependent design decisions rather than implementation details.

What carries the argument

The organizing device is a two-axis taxonomy: spatiotemporal granularity (static 3D, dynamic 3D, 4D) on one axis, and representation type (vision-based, NeRF-based, Gaussian-splatting-based) on the other. Physically, the load-bearing machinery is the Material Point Method, the Finite Element Method, and the DiffTaichi differentiable-programming framework, which supply the simulation steps that generative models optimize against. Quantitatively, the instrument is Table 3, an error comparison of four dynamic-3D methods on the Synthetic PAC-NeRF dataset, which carries the survey's strongest empirical claim about which approach is most robust.

What would settle it

Re-run all four methods from their released code on the Synthetic PAC-NeRF dataset using one shared evaluation script and the same seeds; if the per-material absolute errors and the overall ranking change materially, the survey's benchmark conclusion fails. A lighter check is to compare the transcribed numbers in Table 3 with the original papers' reported tables.

Watch

Extended reading notes

Core claim

The paper's central claim is that physics-grounded generative models are now systematic enough to survey, and that the right organizing principle is spatiotemporal granularity crossed with representation type. Static 3D generation enforces equilibrium and structural plausibility; dynamic 3D generation adds short-term motion via differentiable simulation; 4D generation targets long-range spatiotemporal consistency. On the Synthetic PAC-NeRF benchmark, the paper reports that the Gaussian-splatting method Unleashing achieves the lowest average physical-parameter errors for most Newtonian fluids, non-Newtonian fluids, and plasticine, while GIC leads on elastic materials and LPO is best on some yield-stress parameters; PAC-NeRF shows the largest errors. The intended conclusion is that integrating differentiable physics (MPM/FEM) into generative models improves physical plausibility, and that different representations carry different material-specific strengths.

Load-bearing premise

The load-bearing premise is that the Table 3 errors for PAC-NeRF, LPO, GIC, and Unleashing were produced under identical, directly comparable evaluation protocols on the same Synthetic PAC-NeRF data; if the original papers used different preprocessing, seeds, or hyperparameters, the survey's ranking and its 'Unleashing most robust' conclusion are not supported.

Editorial extensions

If this is right

Static 3D pipelines can be pushed from visual fidelity toward structural integrity: standability, equilibrium, and physical compatibility of multi-part objects become testable generation objectives.
In dynamic 3D, the differentiable solver is the main bridge, and the reported trade-offs imply that elastic materials favor Gaussian-splatting methods while fluid and plasticine materials also favor splatting, with NeRF-based particle methods still competitive for geometry-agnostic identification.
4D generation does not need diffusion models: simulating a Gaussian field with MPM can produce temporally consistent content at much lower cost, as the survey's Phy124 discussion shows.
Benchmarking should report per-material physical-parameter errors rather than a single average, because the comparative analysis shows material-dependent winners.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The Table 3 ranking is provisional: the four systems were tuned and reported by different groups, so a single-codebase rerun with shared seeds and preprocessing could change the ordering; this is an inference, not a claim the survey makes.
The taxonomy predicts a transfer effect: methods that optimize physical parameters over an entire sequence (LPO, Unleashing, GIC) should generalize better to unseen soft-body and fluid scenes than methods anchored to a first frame (PAC-NeRF).
As large language models improve at material inference, the field's bottleneck may shift from estimating parameters to choosing the right constitutive model; hybrid learned-plus-simulated solvers may then outperform either pure physics or pure learning.
The survey's dataset inventory could seed a standardized physics-AIGC benchmark that controls for solver, initializer, and seed, which would let future work isolate the contribution of physics priors from the contribution of representation.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The survey's taxonomy is genuinely useful: organizing physics-aware generation into static 3D, dynamic 3D (vision/NeRF/GS-based), and 4D, with a background section on MPM/FEM/DiffTaichi, gives newcomers a decent map of a fast-moving field. The curated list is also a good resource. That part deserves credit.

The soft spot is Table 3, and it's not minor. The authors conclude that “Unleashing demonstrates the most robust overall performance” based on four numbers rows, but nowhere do they say where those numbers came from, what protocol was used, or which checkpoints/seeds were involved. The row for “Unleashing” cites Liu et al. 2025b, a CVPR paper on 4D physical scene simulation with video diffusion; that paper does not obviously report per-object physical parameter errors (Δμ, Δκ, ΔτY, ΔE, Δν, Δθfric) on the Synthetic PAC-NeRF dataset. If the values were not taken directly from the cited papers or from a clearly described reimplementation, the ranking is unsupported. Even if they were, the survey doesn't show the four methods were evaluated under identical conditions, so the comparative claim is load-bearing but unverified. There's also a formatting problem: entries like “0.890.890.89” suggest the table has been mangled in production, making it illegible.

There's also a factual misattribution: the text says “Gaussian Splashing also provides a novel underwater dataset TableDB,” but that dataset is from Mualem et al. 2024, not from the Gaussian Splashing paper by Feng et al. 2025. That's a concrete error a referee would catch. Minor prose issues (typos like “simulayion”, “expends”, “K-Plances”, “PPM” for Phys4DGen) are easy to fix.

I disagree with the reader on one point: the taxonomy itself is not undermined by the benchmark problems. As a survey, its value is synthesis and guidance, and that part holds up. But the quantitative comparison is presented as a key contribution, so the authors need to either provide a verifiable provenance trail for Table 3 or remove it entirely and rely on the taxonomy.

Who is this for? Researchers entering physics-aware AIGC who want an overview and a reading list. It deserves a serious referee—the topic is timely and the taxonomy is a reasonable organizing scheme—but the paper should go back for major revision on the benchmark section. If the authors can't back up Table 3, I'd recommend cutting it rather than shipping numbers that can't be checked.

Referee Report

4 major / 5 minor

Summary. This manuscript is a survey of physics-aware 3D and 4D generation. It proposes a taxonomy with three categories (static 3D, dynamic 3D, and 4D generation), organizes dynamic 3D methods by representation (vision-based, NeRF-based, and Gaussian-Splatting-based), describes relevant datasets and metrics, and presents a quantitative comparison of four methods (PAC-NeRF, LPO, GIC, and Unleashing) on Synthetic PAC-NeRF. The central comparative claim is that Unleashing shows the most robust overall performance across fluid, plasticine, elasticity, and sand materials.

Significance. The paper's organizational contribution is genuinely useful: Figure 1 and the accompanying taxonomy assemble a recent and fast-moving literature, Table 2 summarizes material parameters compactly, and the curated GitHub list is a practical resource. If the quantitative comparison were fully substantiated, the survey would be a valuable entry point for researchers choosing among physics-aware generative methods. At present, however, the load-bearing comparison is not yet supported: the provenance and comparability of the Table 3 numbers are unclear, and one of the four methods is classified inconsistently. The significance is therefore mostly structural and bibliographic until the comparison is repaired.

major comments (4)

[Section 4, Table 3] The conclusion that 'Unleashing demonstrates the most robust overall performance' rests entirely on Table 3, but the manuscript does not state where the numerical values come from, which checkpoints or hyperparameters were used, or whether all four methods were evaluated under an identical protocol (same viewpoints, same MLS-MPM reference, same error definition, same preprocessing) on the Synthetic PAC-NeRF dataset. The cited Unleashing paper (Liu et al., 2025b) is described elsewhere in the survey as a 4D simulation framework, and its public materials do not obviously report per-object physical-parameter MAEs (Δμ, Δκ, ΔτY, Δη, ΔE, Δν, Δθfric) for the nine Synthetic PAC-NeRF instances. If the authors reimplemented the method or obtained unpublished numbers, that must be disclosed; otherwise the ranking in Table 3 is unsupported. Please add a reproducibility appendix giving the exact source or protocol for every value in the table.
[Section 3.3 vs. Section 4] There is an internal inconsistency in the taxonomy: Unleashing (Liu et al., 2025b) is placed under '4D Generation' in Figure 1 and discussed in Section 3.3 as a 4D simulation framework, yet Section 4 groups it with 'GS-based Methods' for dynamic 3D generation and compares it on Synthetic PAC-NeRF. Since the proposed taxonomy is one of the manuscript's main contributions, the authors should either reconcile the classification or explicitly motivate why a 4D method is benchmarked in a dynamic 3D setting.
[Section 3.2, GS-based paragraph] The sentence 'Additionally, Gaussian Splashing also provides a novel underwater dataset TableDB' is ambiguous and likely misattributes the dataset. The immediately preceding sentences discuss two distinct works: 'Gaussian Splashing (GSP) [Feng et al., 2025]' for fluid-solid coupling and 'Gaussian Splashing for underwater imagery [Mualem et al., 2024]' for underwater rendering. As written, the unqualified name 'Gaussian Splashing' appears to credit Feng et al. with the TableDB dataset, whereas the dataset belongs to the underwater work of Mualem et al. Please correct the attribution and use distinct names for the two works throughout.
[Section 1 and Section 4] The abstract promises a 'systematic' review, and Section 4 claims to compare 'state-of-the-art approaches', but the survey gives no search strategy, inclusion criteria, or rationale for selecting exactly PAC-NeRF, LPO, GIC, and Unleashing for the benchmark. Without this information, the reader cannot assess whether the taxonomy and the comparative analysis are complete or biased toward a particular subgroup of methods. Please state the paper collection procedure and the selection criteria for the four benchmarked methods, or soften the systematicity claim.

minor comments (5)

[Table 3] The table is not reliably readable: values appear concatenated (e.g., '0.890.890.89', '3.0×1033.0×1033.0×103'), and the column boundaries between methods are unclear. Please re-typeset the table with clearly separated per-method columns and unambiguous cell entries.
[Section 4.2] The metrics discussion introduces CLIP score, Semantic Adherence, Physical Commonsense, PSNR, SSIM, and MAE, but the only comparison actually reported in Table 3 is MAE of physical parameters. Please either report the additional metrics or state explicitly that they are discussed for context but not used in the benchmark.
[Throughout] There are several typographical issues: 'K-Plances' should be 'K-Planes' (Section 2.3), 'simulayion' should be 'simulation' (Section 2.4), 'sheer modulus' should be 'shear modulus' (Table 2 context), 'exsiting' should be 'existing' (Section 3.2), and the capitalization of 'unleashing' is inconsistent with 'Unleashing'.
[References] The reference [OpenAI, 2023] is the GPT-4V system card, but Section 3.3 mentions GPT-4o when describing Phys4DGen. Please update the citation to the correct model card or cite the specific GPT-4o release.
[Section 4.1] The description of Synthetic PAC-NeRF states that each scene is 'captured from 11 viewpoints' with ground truth generated by MLS-MPM; please verify these details against the original PAC-NeRF paper and cite the exact dataset configuration, since the subsequent comparison depends on this dataset description.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the survey contains no derivation of its own, and its comparative claims are external evidence, not fitted inputs or self-citations.

full rationale

This paper is a survey of physics-aware 3D and 4D generation methods. It makes no predictions, fits no parameters, and derives no quantitative results of its own. The taxonomy in Figure 1 and the qualitative descriptions in Sections 3 and 4 are organizing summaries of external works, with citations to the original methods. Table 3 transcribes reported physical-parameter errors from PAC-NeRF, LPO, GIC, and Unleashing; even if the provenance or comparability of those numbers were questionable, that would be an issue of evidence quality and not circularity, because the survey does not define any method in terms of the compared outcomes or derive the ranking from its own assumptions. The only self-references are institutional and the companion GitHub list, which are not load-bearing. No step reduces to its own input, no fitted parameter is renamed as a prediction, and no uniqueness or ansatz is imported from the authors' prior work. The central claim, that physics priors are integrated into 3D and 4D generation in identifiable categories, is supported by citation to independent external sources rather than by circular construction. Therefore the circularity score is 0.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The survey introduces no free parameters or invented entities. Its central claims rest on domain assumptions about the completeness of the taxonomy, the accuracy of the method summaries, and the comparability of the reproduced benchmark table.

assumptions (3)

domain assumption The taxonomy categories (static 3D, dynamic 3D, and 4D generation) are exhaustive and mutually exclusive for the surveyed literature.
The paper places each method into exactly one category in Figure 1 without defining operational boundary rules, e.g., the distinction between dynamic 3D and 4D is described only as 'short-term temporal dynamics' versus 'continuous spatiotemporal synthesis' (Section 3).
domain assumption The compared methods in Table 3 are evaluated under compatible settings on the Synthetic PAC-NeRF dataset.
The survey does not report the evaluation protocol, seeds, or hyperparameters for the four methods, so the comparability of the numbers is assumed (Section 4, Table 3).
domain assumption The described method summaries in Sections 3.1 to 3.3 accurately reflect the original cited papers.
The survey relies on its own paraphrases, and at least one apparent misattribution occurs in the 'Gaussian Splashing' paragraph, so the accuracy of other summaries is assumed rather than verified.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Grounding Creativity in Physics: A Brief Survey of Physical Priors in AIGC." pith.science (2026). https://pith.science/paper/OVMDWDQ6

@misc{pith2026250207007,
  author       = {Pith},
  title        = {Pith review of: Grounding Creativity in Physics: A Brief Survey of Physical Priors in AIGC},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OVMDWDQ6}},
  note         = {Machine review of arXiv:2502.07007}
}

read the original abstract

Recent advancements in AI-generated content have significantly improved the realism of 3D and 4D generation. However, most existing methods prioritize appearance consistency while neglecting underlying physical principles, leading to artifacts such as unrealistic deformations, unstable dynamics, and implausible objects interactions. Incorporating physics priors into generative models has become a crucial research direction to enhance structural integrity and motion realism. This survey provides a review of physics-aware generative methods, systematically analyzing how physical constraints are integrated into 3D and 4D generation. First, we examine recent works in incorporating physical priors into static and dynamic 3D generation, categorizing methods based on representation types, including vision-based, NeRF-based, and Gaussian Splatting-based approaches. Second, we explore emerging techniques in 4D generation, focusing on methods that model temporal dynamics with physical simulations. Finally, we conduct a comparative analysis of major methods, highlighting their strengths, limitations, and suitability for different materials and motion dynamics. By presenting an in-depth analysis of physics-grounded AIGC, this survey aims to bridge the gap between generative models and physical realism, providing insights that inspire future research in physically consistent content generation.

Figures

Figures reproduced from arXiv: 2502.07007 by the authors.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

Canvas3D: Empowering Precise Spatial Control for Image Generation with Constraints from a 3D Virtual Canvas
cs.HC 2025-08 conditional novelty 6.0 of 10

Canvas3D lets users arrange objects in a 3D canvas generated from a text prompt, then feeds depth, skeleton, and lighting constraints to diffusion models to produce images that match the layout.

Reference graph

Works this paper leans on

63 extracted references · 48 canonical work pages · cited by 1 Pith paper

[1]

ParticleNeRF: A particle-based encoding for online neural radiance fields

[Abou-Chakra et al., 2024] Jad Abou-Chakra, Feras Dayoub, and Niko S¨underhauf. ParticleNeRF: A particle-based encoding for online neural radiance fields. In WACV,

work page 2024
[3]

VideoPhy: Evaluating physical commonsense for video generation

[Bansal et al., 2025] Hritik Bansal, Zongyu Lin, Tianyi Xie, Zeshun Zong, Michal Yarom, Yonatan Bitton, Chenfanfu Jiang, Yizhou Sun, Kai-Wei Chang, and Aditya Grover. VideoPhy: Evaluating physical commonsense for video generation. ICLR,

work page 2025
[5]

GIC: Gaussian-informed continuum for physical property identification and simulation

[Cai et al., 2024] Junhao Cai, Yuji Yang, Weihao Yuan, Yisheng He, Zilong Dong, Liefeng Bo, Hui Cheng, and Qifeng Chen. GIC: Gaussian-informed continuum for physical property identification and simulation. NeurIPS,

work page 2024
[6]

A comprehensive survey of ai-generated content (AIGC): A history of generative ai from gan to chatgpt

[Cao et al., 2023] Yihan Cao, Siyu Li, Yixin Liu, Zhiling Yan, Yu- tong Dai, Philip S Yu, and Lichao Sun. A comprehensive survey of ai-generated content (AIGC): A history of generative ai from gan to chatgpt. arXiv preprint arXiv:2303.04226,

arXiv 2023
[8]

Atlas3D: Physically constrained self-supporting text-to-3d for simulation and fabrication

[Chen et al., 2024] Yunuo Chen, Tianyi Xie, Zeshun Zong, Xuan Li, Feng Gao, Yin Yang, Ying Nian Wu, and Chenfanfu Jiang. Atlas3D: Physically constrained self-supporting text-to-3d for simulation and fabrication. NeurIPS,

work page 2024
[9]

Neural point cloud rendering via multi-plane projection

[Dai et al., 2020] Peng Dai, Yinda Zhang, Zhuwen Li, Shuaicheng Liu, and Bing Zeng. Neural point cloud rendering via multi-plane projection. In CVPR,

work page 2020
[12]

Gaussian Splashing: Uni- fied particles for versatile motion synthesis and rendering

[Feng et al., 2025] Yutao Feng, Xiang Feng, Yintong Shang, Ying Jiang, Chang Yu, Zeshun Zong, Tianjia Shao, Hongzhi Wu, Kun Zhou, Chenfanfu Jiang, and Yin Yang. Gaussian Splashing: Uni- fied particles for versatile motion synthesis and rendering. CVPR,

work page 2025
[13]

K-Planes: Explicit radiance fields in space, time, and appearance

[Fridovich-Keil et al., 2023] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-Planes: Explicit radiance fields in space, time, and appearance. In CVPR,

work page 2023

Show all 63 references

[14]

Dynamic view synthesis from dynamic monocular video

[Gao et al., 2021] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia- Bin Huang. Dynamic view synthesis from dynamic monocular video. In ICCV,

2021
[17]

De- noising diffusion probabilistic models

[Ho et al., 2020] Jonathan Ho, Ajay Jain, and Pieter Abbeel. De- noising diffusion probabilistic models. NeurIPS,

2020
[18]

A moving least squares material point method with displacement discontinuity and two-way rigid body coupling

[Hu et al., 2018] Yuanming Hu, Yu Fang, Ziheng Ge, Ziyin Qu, Yixin Zhu, Andre Pradhana, and Chenfanfu Jiang. A moving least squares material point method with displacement discontinuity and two-way rigid body coupling. TOG,

2018
[20]

DreamPhysics: Learning physical properties of dynamic 3d gaussians with video diffusion priors

[Huang et al., 2025] Tianyu Huang, Yihan Zeng, Hui Li, Wangmeng Zuo, and Rynson WH Lau. DreamPhysics: Learning physical properties of dynamic 3d gaussians with video diffusion priors. AAAI,

2025
[21]

The material point method for simulating continuum materials

[Jiang et al., 2016] Chenfanfu Jiang, Craig Schroeder, Joseph Teran, Alexey Stomakhin, and Andrew Selle. The material point method for simulating continuum materials. In SIGGRAPH

2016
[23]

3d gaussian splatting for real-time radiance field rendering

[Kerbl et al., 2023] Bernhard Kerbl, Georgios Kopanas, Thomas Leimk¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. TOG,

2023
[24]

Differentiable physics simulation of dynamics-augmented neural objects

[Le Cleac’h et al., 2023] Simon Le Cleac’h, Hong-Xing Yu, Michelle Guo, Taylor Howell, Ruohan Gao, Jiajun Wu, Zachary Manchester, and Mac Schwager. Differentiable physics simulation of dynamics-augmented neural objects. RAL,

2023
[25]

PAC-NeRF: Physics augmented continuum neural radiance fields for geometry-agnostic system identification

[Li et al., 2023] Xuan Li, Yi-Ling Qiao, Peter Yichen Chen, Kr- ishna Murthy Jatavallabhula, Ming Lin, Chenfanfu Jiang, and Chuang Gan. PAC-NeRF: Physics augmented continuum neural radiance fields for geometry-agnostic system identification. In ICLR,

2023
[26]

Generative Image Dynamics

[Li et al., 2024] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative Image Dynamics. In CVPR,

2024
[27]

PhiP-G: Physics-guided text-to-3d compositional scene generation

[Li et al., 2025a] Qixuan Li, Chao Wang, Zongjin He, and Yan Peng. PhiP-G: Physics-guided text-to-3d compositional scene generation. arXiv preprint arXiv:2502.00708,

arXiv
[28]

Phy124: Fast physics-driven 4d content generation from a single image

[Lin et al., 2024a] Jiajing Lin, Zhenzhong Wang, Yongjie Hou, Yuzhou Tang, and Min Jiang. Phy124: Fast physics-driven 4d content generation from a single image. arXiv preprint arXiv:2409.07179,

arXiv
[29]

Phys4DGen: A physics-driven framework for controllable and efficient 4d content generation from a single image

[Lin et al., 2024b] Jiajing Lin, Zhenzhong Wang, Shu Jiang, Yongjie Hou, and Min Jiang. Phys4DGen: A physics-driven framework for controllable and efficient 4d content generation from a single image. arXiv preprint arXiv:2411.16800,
[30]

OMNIPHYSGS: 3d constitutive gaussians for general physics-based dynamics generation

[Lin et al., 2025] Yuchen Lin, Chenguo Lin, Jianjin Xu, and Yadong Mu. OMNIPHYSGS: 3d constitutive gaussians for general physics-based dynamics generation. ICLR,

2025
[31]

Physics3D: Learning physical properties of 3d gaussians via video diffusion

[Liu et al., 2024a] Fangfu Liu, Hanyang Wang, Shunyu Yao, Shengjun Zhang, Jie Zhou, and Yueqi Duan. Physics3D: Learning physical properties of 3d gaussians via video diffusion. arXiv preprint arXiv:2406.04338,

arXiv
[32]

A comprehensive survey on 3d content generation

[Liu et al., 2024b] Jian Liu, Xiaoshui Huang, Tianyu Huang, Lu Chen, Yuenan Hou, Shixiang Tang, Ziwei Liu, Wanli Ouyang, Wangmeng Zuo, Junjun Jiang, et al. A comprehensive survey on 3d content generation. arXiv preprint arXiv:2402.01166,

arXiv
[33]

Generative phys- ical ai in vision: A survey

[Liu et al., 2025a] Daochang Liu, Junyu Zhang, Anh-Dung Dinh, Eunbyung Park, Shichao Zhang, and Chang Xu. Generative phys- ical ai in vision: A survey. arXiv preprint arXiv:2501.10928 ,

arXiv
[34]

Unified simulation of elastic rods, shells, and solids

[Martin et al., 2010] Sebastian Martin, Peter Kaufmann, Mario Botsch, Eitan Grinspun, and Markus Gross. Unified simulation of elastic rods, shells, and solids. TOG,

2010
[36]

Gaussian splashing: Direct volumetric render- ing underwater

[Mualem et al., 2024] Nir Mualem, Roy Amoyal, Oren Freifeld, and Derya Akkaynak. Gaussian splashing: Direct volumetric render- ing underwater. arXiv preprint arXiv:2411.19588,

2024
[37]

Instant neural graphics primitives with a multiresolution hash encoding

[M¨uller et al., 2022] Thomas M¨uller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. TOG,

2022
[38]

GPT-4V(ision) System Card,

[OpenAI, 2023] OpenAI. GPT-4V(ision) System Card,

2023
[39]

[Park et al., 2021] Keunhong Park, Utkarsh Sinha, Jonathan T Bar- ron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla

https: //openai.com/index/gpt-4o-system-card/. [Park et al., 2021] Keunhong Park, Utkarsh Sinha, Jonathan T Bar- ron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV,

2021
[40]

Barron, and Ben Mildenhall

[Poole et al., 2023] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3d using 2d diffusion. ICLR,

2023
[41]

D-NeRF: Neural radi- ance fields for dynamic scenes

[Pumarola et al., 2021] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural radi- ance fields for dynamic scenes. In CVPR,

2021
[42]

Learning transferable visual models from natural language supervision

[Radford et al., 2021] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML,

2021
[43]

MotionCraft: Physics-based zero-shot video generation

[Savant Aira et al., 2024] Luca Savant Aira, Antonio Montanaro, Emanuele Aiello, Diego Valsesia, and Enrico Magli. MotionCraft: Physics-based zero-shot video generation. NeurIPS,

2024
[44]

Diffusion-based signed distance fields for 3d shape generation

[Shim et al., 2023] Jaehyeok Shim, Changwoo Kang, and Kyung- don Joo. Diffusion-based signed distance fields for 3d shape generation. In CVPR,

2023
[45]

PhysMotion: Physics-grounded dynamics from a single image

[Tan et al., 2024] Xiyang Tan, Ying Jiang, Xuan Li, Zeshun Zong, Tianyi Xie, Yin Yang, and Chenfanfu Jiang. PhysMotion: Physics-grounded dynamics from a single image. arXiv preprint arXiv:2411.17189,

2024 arXiv
[46]

DreamGaussian: Generative gaussian splat- ting for efficient 3d content creation

[Tang et al., 2024] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. DreamGaussian: Generative gaussian splat- ting for efficient 3d content creation. ICLR,

2024
[47]

GaMeS: Mesh-based adapting and modification of gaussian splatting

[Waczy´nska et al., 2024] Joanna Waczy ´nska, Piotr Borycki, Sławomir Tadeja, Jacek Tabor, and Przemysław Spurek. GaMeS: Mesh-based adapting and modification of gaussian splatting. arXiv preprint arXiv:2402.01459,

2024 arXiv
[48]

Compositional 4d dy- namic scenes understanding with physics priors for video question answering

[Wang et al., 2025] Xingrui Wang, Wufei Ma, Angtian Wang, Shuo Chen, Adam Kortylewski, and Alan Yuille. Compositional 4d dy- namic scenes understanding with physics priors for video question answering. ICLR,

2025
[49]

SynSin: End-to-end view synthesis from a single image

[Wiles et al., 2020] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. SynSin: End-to-end view synthesis from a single image. In CVPR,

2020
[50]

Physics 101: Learning physical object properties from unlabeled videos

[Wu et al., 2016] Jiajun Wu, Joseph J Lim, Hongyi Zhang, Joshua B Tenenbaum, and William T Freeman. Physics 101: Learning physical object properties from unlabeled videos. In BMVC,

2016
[51]

Omniobject3D: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation

[Wu et al., 2023] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3D: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In CVPR,

2023
[52]

4d gaussian splatting for real-time dynamic scene rendering

[Wu et al., 2024] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In CVPR,

2024
[53]

PhysGaussian: Physics-integrated 3d gaussians for generative dynamics

[Xie et al., 2024] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. PhysGaussian: Physics-integrated 3d gaussians for generative dynamics. In CVPR,

2024
[54]

Precise-physics driven text-to-3d generation

[Xu et al., 2024] Qingshan Xu, Jiao Liu, Melvin Wong, Caishun Chen, and Yew-Soon Ong. Precise-physics driven text-to-3d generation. arXiv preprint arXiv:2403.12438,

2024 arXiv
[55]

PhyT2V: LLM-guided iterative self-refinement for physics- grounded text-to-video generation

[Xue et al., 2025] Qiyao Xue, Xiangyu Yin, Boyuan Yang, and Wei Gao. PhyT2V: LLM-guided iterative self-refinement for physics- grounded text-to-video generation. CVPR,

2025
[56]

PhyCAGE: Physically plausible compositional 3D asset generation from a single image

[Yan et al., 2024] Han Yan, Mingrui Zhang, Yang Li, Chao Ma, and Pan Ji. PhyCAGE: Physically plausible compositional 3D asset generation from a single image. NeurIPS,

2024
[57]

Real-time photorealistic dynamic scene representa- tion and rendering with 4d gaussian splatting

[Yang et al., 2024] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representa- tion and rendering with 4d gaussian splatting. ICLR,

2024
[58]

GaussianDreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models

[Yi et al., 2024] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xing- gang Wang. GaussianDreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. InCVPR,

2024
[59]

Trans4D: Realistic geometry-aware transition for compositional text-to-4d synthesis

[Zeng et al., 2024] Bohan Zeng, Ling Yang, Siyu Li, Jiaming Liu, Zixiang Zhang, Juanxi Tian, Kaixin Zhu, Yongzhen Guo, Fu- Yun Wang, Minkai Xu, et al. Trans4D: Realistic geometry-aware transition for compositional text-to-4d synthesis. arXiv preprint arXiv:2410.07155,

2024 arXiv
[60]

MEGA: Memory-efficient 4d gaussian splatting for dynamic scenes

[Zhang et al., 2024b] Xinjie Zhang, Zhening Liu, Yifan Zhang, Xingtong Ge, Dailan He, Tongda Xu, Yan Wang, Zehong Lin, Shuicheng Yan, and Jun Zhang. MEGA: Memory-efficient 4d gaussian splatting for dynamic scenes. arXiv preprint arXiv:2410.13613,

arXiv
[61]

Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians

[Zhong et al., 2024] Licheng Zhong, Hong-Xing Yu, Jiajun Wu, and Yunzhu Li. Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians. ECCV,

2024
[62]

LAYOUTDREAMER: Physics-guided layout for text-to-3d compositional scene generation

[Zhou et al., 2025] Yang Zhou, Zongjin He, Qixuan Li, and Chao Wang. LAYOUTDREAMER: Physics-guided layout for text-to-3d compositional scene generation. arXiv preprint arXiv:2502.01949,

2025
[63]

The finite element method: solid mechanics

[Zienkiewicz and Taylor, 2000] Olgierd Cecil Zienkiewicz and Robert Leroy Taylor. The finite element method: solid mechanics. Butterworth-heinemann, 2000

2000
[2010]

Nerf: Representing scenes as neural radiance fields for view synthesis

[Mildenhall et al., 2020] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. ECCV,

2020
[2014]

Physically compatible 3d object modeling from a single image

[Guo et al., 2024] Minghao Guo, Bohan Wang, Pingchuan Ma, Tianyuan Zhang, Crystal Elaine Owens, Chuang Gan, Joshua B Tenenbaum, Kaiming He, and Wojciech Matusik. Physically compatible 3d object modeling from a single image. NeurIPS,

2024
[2016]

Improving physics-augmented continuum neural radiance field-based geometry-agnostic system identification with lagrangian particle optimization

[Kaneko, 2024] Takuhiro Kaneko. Improving physics-augmented continuum neural radiance field-based geometry-agnostic system identification with lagrangian particle optimization. In CVPR,

2024
[2018]

DiffTaichi: Differentiable programming for physical simulation

[Hu et al., 2020] Yuanming Hu, Luke Anderson, Tzu-Mao Li, Qi Sun, Nathan Carr, Jonathan Ragan-Kelley, and Fredo Durand. DiffTaichi: Differentiable programming for physical simulation. In ICLR,

2020
[2020]

Fast dynamic radiance fields with time-aware neural voxels

[Fang et al., 2022] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH,

2022
[2021]

Generative adversarial nets

[Goodfellow et al., 2014] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS,

2014
[2022]

PIE-NeRF: Physics-based interactive elastodynamics with nerf

[Feng et al., 2024] Yutao Feng, Yintong Shang, Xuan Li, Tianjia Shao, Chenfanfu Jiang, and Yin Yang. PIE-NeRF: Physics-based interactive elastodynamics with nerf. In CVPR,

2024
[2023]

NeuMA: Neural material adaptor for visual grounding of intrinsic dynamics

[Cao et al., 2024] Junyi Cao, Shanyan Guan, Yanhao Ge, Wei Li, Xiaokang Yang, and Chao Ma. NeuMA: Neural material adaptor for visual grounding of intrinsic dynamics. NeurIPS,

2024
[2024]

Physics-informed computer vision: A review and perspectives

[Banerjee et al., 2024] Chayan Banerjee, Kien Nguyen, Clinton Fookes, and Karniadakis George. Physics-informed computer vision: A review and perspectives. ACM COMPUT SURV,

2024
[2025]

GASP: Gaussian splatting for physic-based simulations

[Borycki et al., 2024] Piotr Borycki, Weronika Smolak, Joanna Waczy´nska, Marcin Mazur, Sławomir Tadeja, and Przemysław Spurek. GASP: Gaussian splatting for physic-based simulations. arXiv preprint arXiv:2409.05819,

2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

ParticleNeRF: A particle-based encoding for online neural radiance fields

[Abou-Chakra et al., 2024] Jad Abou-Chakra, Feras Dayoub, and Niko S¨underhauf. ParticleNeRF: A particle-based encoding for online neural radiance fields. In WACV,

work page 2024

[2] [3]

VideoPhy: Evaluating physical commonsense for video generation

[Bansal et al., 2025] Hritik Bansal, Zongyu Lin, Tianyi Xie, Zeshun Zong, Michal Yarom, Yonatan Bitton, Chenfanfu Jiang, Yizhou Sun, Kai-Wei Chang, and Aditya Grover. VideoPhy: Evaluating physical commonsense for video generation. ICLR,

work page 2025

[3] [5]

GIC: Gaussian-informed continuum for physical property identification and simulation

[Cai et al., 2024] Junhao Cai, Yuji Yang, Weihao Yuan, Yisheng He, Zilong Dong, Liefeng Bo, Hui Cheng, and Qifeng Chen. GIC: Gaussian-informed continuum for physical property identification and simulation. NeurIPS,

work page 2024

[4] [6]

A comprehensive survey of ai-generated content (AIGC): A history of generative ai from gan to chatgpt

[Cao et al., 2023] Yihan Cao, Siyu Li, Yixin Liu, Zhiling Yan, Yu- tong Dai, Philip S Yu, and Lichao Sun. A comprehensive survey of ai-generated content (AIGC): A history of generative ai from gan to chatgpt. arXiv preprint arXiv:2303.04226,

arXiv 2023

[5] [8]

Atlas3D: Physically constrained self-supporting text-to-3d for simulation and fabrication

[Chen et al., 2024] Yunuo Chen, Tianyi Xie, Zeshun Zong, Xuan Li, Feng Gao, Yin Yang, Ying Nian Wu, and Chenfanfu Jiang. Atlas3D: Physically constrained self-supporting text-to-3d for simulation and fabrication. NeurIPS,

work page 2024

[6] [9]

Neural point cloud rendering via multi-plane projection

[Dai et al., 2020] Peng Dai, Yinda Zhang, Zhuwen Li, Shuaicheng Liu, and Bing Zeng. Neural point cloud rendering via multi-plane projection. In CVPR,

work page 2020

[7] [12]

Gaussian Splashing: Uni- fied particles for versatile motion synthesis and rendering

[Feng et al., 2025] Yutao Feng, Xiang Feng, Yintong Shang, Ying Jiang, Chang Yu, Zeshun Zong, Tianjia Shao, Hongzhi Wu, Kun Zhou, Chenfanfu Jiang, and Yin Yang. Gaussian Splashing: Uni- fied particles for versatile motion synthesis and rendering. CVPR,

work page 2025

[8] [13]

K-Planes: Explicit radiance fields in space, time, and appearance

[Fridovich-Keil et al., 2023] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-Planes: Explicit radiance fields in space, time, and appearance. In CVPR,

work page 2023