Pith. sign in

REVIEW 3 major objections 6 minor 51 references

SonicGauss: Position-Aware Physical Sound Synthesis for 3D Gaussian Representations

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read SonicGauss claims that a 3D Gaussian Splatting representation carries enough geometric and material information to synthesize realistic impact sounds that change with impact location, and that it is the first method to do this directly…

desk verdict A plausible first step toward 3DGS-conditioned impact sound synthesis, but the generalization claim is unsupported until the authors rule out per-object memorization with a real held-out evaluation. read the letter →

arxiv 2507.19835 v1 pith:DLM5DFQH submitted 2025-07-26 cs.SD cs.MM

Chunshi Wang , Hongxing Li , Yawei Luo This is my paper
classification cs.SDcs.MM
keywords 3DGaussianSplattingimpactsoundsynthesisposition-awareaudiodiffusionmodelsPointTransformercontrastivelearningcross-modalgenerationObjectFolder
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

This paper tries to establish that a 3D Gaussian Splatting (3DGS) object representation—the same set of colored ellipsoids used for real-time visual rendering—carries enough geometric and material information to synthesize realistic impact sounds, and that the sound should change with where the object is struck. The authors claim SonicGauss is the first framework to do this directly from 3DGS, without any video, image, or text input at inference time. The payoff would be interactive 3D scenes in virtual reality, games, and digital twins gaining a physically suggestive audio layer by reusing a representation they already have. The paper reports three-stage training on synthetic and real impact-sound datasets, with human listeners rating material, sound quality, and position matching around 0.84–0.91 on a 0–1 scale.

What carries the argument

The machinery is a three-stage transfer pipeline. Stage 1 fine-tunes a pretrained text-to-audio flow-matching diffusion model on impact sounds using material captions generated by a vision-language model, so the model learns the sound of descriptions like 'metal, hollow, resonant.' Stage 2.1 aligns a Point Transformer V3 encoder, which treats Gaussian ellipsoids as multi-channel point clouds, to text embeddings with an InfoNCE contrastive loss; Stage 2.2 then replaces the text encoder with the Gaussian encoder and coarsely fine-tunes the full model on the synthetic dataset. Stage 3 injects the impact position through a sinusoidal frequency encoding and an MLP, merges it with Gaussian features via cross-attention, and fine-tunes the whole system on real multi-position recordings. Each stage hands the next a different capability: material vocabulary, direct geometry-to-sound conditioning, and position sensitivity.

What would settle it

A decisive test is a control that receives only an object-identity tag instead of the Gaussian features, or two 3DGS reconstructions of the same geometry with altered material appearance; if either control matches SonicGauss's reported FAD around 0.73 on ObjectFolder-Real, then the acoustic fidelity is not carried by the ellipsoid parameters.

Watch

Extended reading notes

Core claim

The central claim is that anisotropic Gaussian ellipsoids—each parameterized by position, scale, rotation, opacity, and spherical-harmonic color—encode natural hints about both shape and substance, and a network can read those hints to predict acoustic behavior. SonicGauss couples a Point Transformer V3-based Gaussian Encoder with a diffusion-based audio generator: material semantics are transferred from text by a contrastive objective, then the text encoder is replaced by the Gaussian encoder and the whole pipeline is fine-tuned on impact sounds. Position awareness is added through a neural-radiance-field-style sinusoidal frequency encoding of the 3D hit coordinate, fused with the Gaussian features by cross-attention, and fine-tuned on real recordings that contain many knock positions per object. The paper's reported numbers show steady improvement across stages—FAD falling from 1.68 in the first stage to 0.73 in the final stage—and ablation studies favor both the contrastive alignment and the cross-attention fusion.

Load-bearing premise

The load-bearing premise is that the Gaussian ellipsoid parameters—where each blob sits, its size, orientation, transparency, and color—encode enough about an object's material and shape for the network to predict its impact sound, and that the measured gains come from those properties rather than the model identifying the object from its one sound per object.

Editorial extensions

If this is right

  • A 3DGS scene can be knocked at any coordinate and produce a distinct, plausible sound, so interactive applications gain position-aware audio without video or text conditioning at inference.
  • Because the Gaussian encoder is trained across 1,000 synthetic and 100 real objects, new objects need only a 3DGS reconstruction to become 'knockable' without per-object audio recordings.
  • The staged recipe shows a pretrained text-to-audio diffusion model can be reconditioned on a non-text sensory representation through contrastive alignment, not just fine-tuned on more text.
  • Acoustic response becomes an output of the 3D Gaussian representation alongside geometry and appearance, which points to other physical properties being liftable the same way.

Reading between the lines

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

  • Beyond the paper: the same staged recipe—caption a physical property, contrastively align a geometry encoder to the caption space, then condition a generator on it—could lift other attributes such as deformability, weight, or temperature from 3DGS.
  • Beyond the paper: because the position condition is a continuous 3D coordinate, the framework could be extended to moving contacts like scraping or rolling by conditioning on a time-varying position path.
  • Beyond the paper: the generator could be inverted into an editing tool—search over hit positions or Gaussian parameters to produce a desired impact sound—turning it into acoustic-aware 3D asset design.
  • Beyond the paper: a testable isolation experiment would randomize material captions in Stage 1 while keeping geometry fixed; if audio tracks the captions, the material channel is real, and if not, most of the effect comes from shape or identity.
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.

Referee Report

3 major / 6 minor

Summary. The paper presents SonicGauss, a three-stage pipeline for synthesizing impact sounds from 3D Gaussian Splatting (3DGS) representations. It fine-tunes a text-to-audio diffusion model on material captions (Stage 1), trains a PointTransformer-based Gaussian encoder via contrastive learning with text (Stage 2.1), performs coarse tuning on the ObjectFolder 2.0 dataset (Stage 2.2), and finally adds a position encoder with cross-attention to generate position-aware impact sounds on ObjectFolder-Real (Stage 3). Experiments report FAD, KL Sigmoid, Inception Score, and human perceptual scores, and claim the first direct 3DGS-to-audio mapping.

Significance. If the generalization claim holds, this is a useful step toward interactive audio for 3DGS scenes, avoiding intermediate video or text at inference time. Strengths of the manuscript include a clearly staged training pipeline, the use of both synthetic and real-world multisensory data, a position-aware cross-attention mechanism, and a human evaluation beyond objective metrics. However, the central claim of cross-category generalization is not yet supported by the evidence because the evaluation does not demonstrate held-out object/category performance and because the training setup admits object-identity memorization as an alternative explanation. The reported quantitative and human scores are therefore internally suggestive but externally uncalibrated.

major comments (3)
  1. [§4.1, §4.5, §4.6] The paper claims cross-category generalization but never defines a train/test split. All quantitative results in Table 1 are aggregate on the full datasets, and the human evaluation (§4.6) refers to 'unseen data from the test set' without specifying how the test set was constructed or whether object instances and categories were held out. The Stage 2.1 contrastive loss (§3.4.1) with batch size 8 can be minimized by instance discrimination, and Stage 3 fine-tuning on ~35 recordings per OF-Real object permits per-object template memorization. The results are therefore consistent with object-identity memorization. Please provide an explicit split by object/category, per-object and per-category FAD, and a control condition that conditions on object identity alone; also report human scores on held-out object categories.
  2. [§4.4, Table 1] The KL Sigmoid metric is defined as 'higher scores imply better performance' (Section 4.4), but Table 1 labels the column 'KL Sig↓' (implying lower is better). The text in Section 4.5 treats the increase from 0.3442 to 0.3930 as an improvement, which is consistent with the definition but contradicts the down-arrow. This makes the metric's direction ambiguous and should be corrected in the table and captions.
  3. [§4.6] The human evaluation measures subjective ratings on material matching, sound quality, and position matching, but it does not include a ground-truth audio reference or a baseline method; without these controls, the high scores (0.84–0.91) are difficult to interpret in absolute terms. Please report participant numbers, sample counts, and statistical tests, and include the original recordings and a simple condition as anchors.
minor comments (6)
  1. [§4.2] The preprocessing section refers to 'more details in the appendix,' but the appendix is not included; the camera trajectory (Eq. 9) and the mesh-to-3DGS reconstruction choices could affect the results, so these details should be summarized in the main text or the appendix made available.
  2. [§3.5.1] Equation (5) uses p=(x,y,z) while Section 3.1 uses p_k for the 3DGS position; the notation is confusing and should be disambiguated.
  3. [§4.4] The definition of 'matched frame size' for FAD is not specified; please clarify the audio feature extractor and window size.
  4. [Table 3, §4.8] The ΔGain row for IS Avg (0.0005) and IS Std (2.60×10⁻⁵) shows extremely small differences; please add confidence intervals or significance tests to show these are not noise.
  5. [Table 4, §4.9] In Table 4, the KL Sig direction is again labeled with a down arrow even though higher is better per Section 4.4; the table captions should be corrected.
  6. [§1] The claim 'first work to explore 3DGS for acoustic synthesis' should be supported by a systematic literature comparison or softened to 'to our knowledge' with a broader related-work search, given the rapid growth of multimodal 3D representation learning.

Circularity Check

0 steps flagged · score 2.0 of 10

No construction-level circularity: SonicGauss is an empirical fitting pipeline with distinct training inputs and evaluation metrics; score 2 reflects minor non-load-bearing self-citations and an under-specified test protocol rather than a demonstrated reduction.

full rationale

SonicGauss makes no first-principles derivation; its chain is a set of learned conditional mappings. Stage 1 fine-tunes TangoFlux with a flow-matching loss on captions and OF-2.0 audio (Eq. 2); Stage 2.1 aligns the Gaussian encoder to text embeddings via InfoNCE (Eq. 3); Stage 2.2 swaps the text condition for the Gaussian condition and fine-tunes (Eq. 4); Stage 3 injects a NeRF-style position encoding and fine-tunes on OF-Real (Eqs. 5-8). No equation defines its target in terms of itself, and the reported FAD/KL/IS metrics are not the training objectives, so there is no definitional equivalence between inputs and predictions. The paper's self-citations (refs 24, 26, 27, 34) are contextual related-work mentions, not load-bearing evidence for the central claim. The genuine weaknesses are evaluation-protocol gaps, not circularity: Section 4.5's Table 1 gives no train/test split; Section 4.6 refers to 'unseen data from the test set' without defining that set; Section 4.2 promises preprocessing details 'in the appendix' that are absent; and Section 4.4 declines comparison with prior impact-sound methods. These gaps mean the generalization claim is not independently verified, but that is an external-validity and correctness concern, not evidence that a prediction reduces by construction. Under the hard-evidence rule, I therefore do not flag any specific circular step.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the availability of object meshes that are converted to 3DGS, on the transferability of TangoFlux, on the quality of GPT-4o captions, and on the unvalidated premise that Gaussian parameters contain material information. No new physical entities are introduced.

free parameters (5)
  • InfoNCE temperature tau = initialized to 0.07, learned
    Controls sharpness of contrastive alignment in Stage 2.1; chosen by convention following CLIP, and affects how Gaussian and text features are aligned.
  • Position frequency bands L = 10
    Number of sinusoidal bands in the position encoding, chosen by hand; determines the sensitivity of position conditioning.
  • 3DGS sample points per object = approximately 10,000
    Density of Gaussian primitives sampled from the mesh surface; affects geometric fidelity and therefore the material cues available to the Gaussian encoder.
  • Orbit camera trajectory = azimuth 15 degree steps, elevation {-45, 0, 45} degrees
    Hand-designed capture trajectory used to render meshes before 3DGS optimization; influences whether material and shape details survive conversion.
  • Training schedule = 80 epochs, batch size 2 (8 for contrastive), AdamW LR 1e-4, cosine schedule
    Hyperparameters for all three stages; chosen by the authors with no reported sweep or sensitivity analysis.
assumptions (5)
  • domain assumption 3D Gaussian ellipsoid parameters encode material properties sufficient to infer impact sound.
    Section 3.1 asserts that 3D Gaussians encode geometric and material attributes and that anisotropic Gaussians give natural hints about substance; no experiment validates this specific assumption or measures how much acoustic information the Gaussian parameters carry.
  • domain assumption The pretrained TangoFlux text-to-audio model can be transferred to isolated impact sounds by fine-tuning on captions.
    Stage 1 assumes transfer learning from general audio to impact sounds succeeds with 80 epochs of caption-audio pairs; no analysis of caption quality or transfer failure modes is given.
  • domain assumption GPT-4o generated material captions from 3D mesh renders are accurate enough to supervise material learning.
    Stage 1 and Stage 2.1 use Vision-LLM captions as ground-truth material descriptions; caption errors would propagate to the contrastive alignment and to the final Gaussian encoder.
  • domain assumption ObjectFolder-Real impact recordings with 3D contact positions and ObjectFolder 2.0 FEM sounds are correctly paired and physically consistent.
    The model relies on dataset pairings without any described verification, cleaning, or handling of recording noise.
  • standard math Standard flow-matching and InfoNCE objectives behave as expected in this setting.
    Equations (2), (3), (4), and (8) use standard objectives; the paper assumes these losses train the network successfully.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SonicGauss: Position-Aware Physical Sound Synthesis for 3D Gaussian Representations." pith.science (2026). https://pith.science/paper/DLM5DFQH

@misc{pith2026250719835,
  author       = {Pith},
  title        = {Pith review of: SonicGauss: Position-Aware Physical Sound Synthesis for 3D Gaussian Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DLM5DFQH}},
  note         = {Machine review of arXiv:2507.19835}
}
read the original abstract

While 3D Gaussian representations (3DGS) have proven effective for modeling the geometry and appearance of objects, their potential for capturing other physical attributes-such as sound-remains largely unexplored. In this paper, we present a novel framework dubbed SonicGauss for synthesizing impact sounds from 3DGS representations by leveraging their inherent geometric and material properties. Specifically, we integrate a diffusion-based sound synthesis model with a PointTransformer-based feature extractor to infer material characteristics and spatial-acoustic correlations directly from Gaussian ellipsoids. Our approach supports spatially varying sound responses conditioned on impact locations and generalizes across a wide range of object categories. Experiments on the ObjectFolder dataset and real-world recordings demonstrate that our method produces realistic, position-aware auditory feedback. The results highlight the framework's robustness and generalization ability, offering a promising step toward bridging 3D visual representations and interactive sound synthesis. Project page: https://chunshi.wang/SonicGauss

Figures

Figures reproduced from arXiv: 2507.19835 by the authors.

Figure 1
Figure 1. SonicGauss: A novel framework for interactive impact sound synthesis from 3D Gaussian Splatting (3DGS). Our [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of SonicGauss for interactive impact [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overview of our SonicGauss framework with its three-stage approach. Stage 1 (T2A Task Transfer): Adapting [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Comparison between ground truth (GT) and SonicGauss-generated spectrograms for impact sounds on various [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Detailed spectrogram analysis comparing ground [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 28 canonical work pages

  1. [1]

    Andrea Agostinelli, Timo I Denk, Zalán Borsos, Jesse Engel, Mauro Verzetti, An- toine Caillon, Qingqing Huang, Aren Jansen, Adam Roberts, Marco Tagliasacchi, et al. 2023. Musiclm: Generating music from text.arXiv preprint arXiv:2301.11325 (2023)

    arXiv 2023

  2. [2]

    Kan Chen, Chuanxi Zhang, Chen Fang, Zhaowen Wang, Trung Bui, and Ram Nevatia. 2018. Visually indicated sound generation by perceptually optimized classification. InProceedings of the European Conference on Computer Vision (ECCV) Workshops. 0–0

    work page 2018

  3. [3]

    Peihao Chen, Yang Zhang, Mingkui Tan, Hongdong Xiao, Deng Huang, and Chuang Gan. 2020. Generating visually aligned sound from videos.IEEE Trans- actions on Image Processing29 (2020), 8292–8302

    work page 2020

  4. [4]

    Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. 2023. Control-a-video: Controllable text-to-video generation with diffusion models.arXiv e-prints(2023), arXiv–2305

    work page 2023

  5. [5]

    Samuel Clarke, Negin Heravi, Mark Rau, Ruohan Gao, Jiajun Wu, Doug James, and Jeannette Bohg. 2022. Diffimpact: Differentiable rendering and identification of impact sounds. InConference on Robot Learning. PMLR, 662–673

    work page 2022

  6. [6]

    Jade Copet, Felix Kreuk, Itai Gat, Tal Remez, David Kant, Gabriel Synnaeve, Yossi Adi, and Alexandre Défossez. 2023. Simple and controllable music generation. Advances in Neural Information Processing Systems36 (2023), 47704–47720

    work page 2023

  7. [7]

    Ruohan Gao, Yiming Dou, Hao Li, Tanmay Agarwal, Jeannette Bohg, Yunzhu Li, Li Fei-Fei, and Jiajun Wu. 2023. The OBJECTFOLDER BENCHMARK: Multisensory Learning with Neural and Real Objects. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 17276–17286

    work page 2023

  8. [8]

    Ruohan Gao, Zilin Si, Yen-Yu Chang, Samuel Clarke, Jeannette Bohg, Li Fei-Fei, Wenzhen Yuan, and Jiajun Wu. 2022. ObjectFolder 2.0: A Multisensory Object Dataset for Sim2Real Transfer. InCVPR

    work page 2022

Show all 51 references
  1. [9]

    Riccardo Fosco Gramaccioni, Christian Marinoni, Emilian Postolache, Marco Comunità, Luca Cosmo, Joshua D Reiss, and Danilo Comminiello. 2024. Stable- V2A: Synthesis of Synchronized Sound Effects with Temporal and Semantic Controls.arXiv preprint arXiv:2412.15023(2024)

    2024 arXiv

  2. [10]

    Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. 2022. Video diffusion models.Advances in Neu- ral Information Processing Systems35 (2022), 8633–8646

    2022

  3. [11]

    Jiun Tian Hoe, Xudong Jiang, Chee Seng Chan, Yap-Peng Tan, and Weipeng Hu

  4. [12]

    Yaosi Hu, Zhenzhong Chen, and Chong Luo. 2023. Lamd: Latent motion diffusion for video generation.arXiv preprint arXiv:2304.11603(2023)

    2023 arXiv

  5. [13]

    Chia-Yu Hung, Navonil Majumder, Zhifeng Kong, Ambuj Mehrish, Amir Zadeh, Chuan Li, Rafael Valle, Bryan Catanzaro, and Soujanya Poria. 2024. TangoFlux: Super Fast and Faithful Text to Audio Generation with Flow Matching and Clap- Ranked Preference Optimization. arXiv:2412.21037...

    2024 arXiv

  6. [14]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. Gpt-4o system card.arXiv preprint arXiv:2410.21276(2024)

    2024 arXiv

  7. [15]

    Xutong Jin, Chenxi Xu, Ruohan Gao, Jiajun Wu, Guoping Wang, and Sheng Li

  8. [16]

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis

  9. [17]

    InACM SIGGRAPH 2024 Conference Papers

    DiffSound: Differentiable Modal Sound Rendering and Inverse Rendering for Diverse Inference Tasks. InACM SIGGRAPH 2024 Conference Papers. 1–12

    2024

  10. [18]

    Felix Kreuk, Gabriel Synnaeve, Adam Polyak, Uriel Singer, Alexandre Défossez, Jade Copet, Devi Parikh, Yaniv Taigman, and Yossi Adi. 2022. Audiogen: Textually guided audio generation.arXiv preprint arXiv:2209.15352(2022)

    2022 arXiv

  11. [19]

    Yeonghyeon Lee, Inmo Yeon, Juhan Nam, and Joon Son Chung. 2024. Voiceldm: Text-to-speech with environmental context. InICASSP 2024-2024 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 12566–12571

    2024

  12. [20]

    Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. 2020. Diffwave: A versatile diffusion model for audio synthesis.arXiv preprint arXiv:2009.09761(2020)

    2020 arXiv

  13. [21]

    Haohe Liu, Yi Yuan, Xubo Liu, Xinhao Mei, Qiuqiang Kong, Qiao Tian, Yuping Wang, Wenwu Wang, Yuxuan Wang, and Mark D Plumbley. 2024. Audioldm 2: Learning holistic audio generation with self-supervised pretraining.IEEE/ACM Transactions on Audio, Speech, and Language Processing(2024)

    2024

  14. [22]

    Xiaojing Liu, Ogulcan Gurelli, Yan Wang, and Joshua Reiss. 2025. Visual-based spatial audio generation system for multi-speaker environments.arXiv preprint arXiv:2502.07538(2025)

    2025 arXiv

  15. [23]

    Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. 2023. Audioldm: Text-to-audio generation with latent diffusion models.arXiv preprint arXiv:2301.12503(2023)

    2023 arXiv

  16. [24]

    Yawei Luo and Yi Yang. 2024. Large language model and domain-specific model collaboration for smart education.Frontiers of Information Technology & Electronic Engineering(2024)

    2024

  17. [25]

    Guoqing Ma, Haoyang Huang, Kun Yan, Liangyu Chen, Nan Duan, Shengming Yin, Changyi Wan, Ranchen Ming, Xiaoniu Song, Xing Chen, et al. 2025. Step- video-t2v technical report: The practice, challenges, and future of video founda- tion model.arXiv preprint arXiv:2502.10248(2025)

    2025 arXiv

  18. [26]

    Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. 2024. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In2024 International Conference on 3D Vision (3DV). IEEE, 800–809

    2024

  19. [27]

    Siwei Meng, Yawei Luo, and Ping Liu. 2025. Grounding creativity in physics: A brief survey of physical priors in aigc. InInternational Joint Conference on Artificial Intelligence

    2025

  20. [28]

    Qiaowei Miao, Kehan Li, Jinsheng Quan, Zhiyuan Min, Shaojie Ma, Yichao Xu, Yi Yang, and Yawei Luo. 2025. Advances in 4d generation: A survey.arXiv preprint arXiv:2503.14501(2025)

    2025 arXiv

  21. [29]

    Shaojie Ma, Yawei Luo, Wei Yang, and Yi Yang. 2025. Mags: Reconstructing and simulating dynamic 3d objects with mesh-adsorbed gaussian splatting. In International Conference on Computer Vision

    2025

  22. [30]

    James F O’Brien, Perry R Cook, and Georg Essl. 2001. Synthesizing sounds from physically based motion. InProceedings of the 28th annual conference on Computer graphics and interactive techniques. 529–536

    2001

  23. [31]

    James F O’Brien, Chen Shen, and Christine M Gatchalian. 2002. Synthesiz- ing sounds from rigid-body simulations. InProceedings of the 2002 ACM SIG- GRAPH/Eurographics symposium on Computer animation. 175–181

    2002

  24. [32]

    Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. 2024. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. InProceedings of the AAAI conference on artificial intelligence, Vol. 38. 4296–4304

    2024

  25. [33]

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. 2023. Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952 (2023)

    2023 arXiv

  26. [34]

    Jinsheng Quan, Chunshi Wang, and Yawei Luo. 2025. ParticleGS: Particle-Based Dynamics Modeling of 3D Gaussians for Prior-free Motion Extrapolation.arXiv preprint arXiv:2505.20270(2025)

    2025

  27. [35]

    Adelson, and William T

    Andrew Owens, Phillip Isola, Josh McDermott, Antonio Torralba, Edward H. Adelson, and William T. Freeman. 2016. Visually Indicated Sounds. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2016

  28. [36]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 10684–10695

    2022

  29. [37]

    Kai Shen, Zeqian Ju, Xu Tan, Yanqing Liu, Yichong Leng, Lei He, Tao Qin, Sheng Zhao, and Jiang Bian. 2023. Naturalspeech 2: Latent diffusion models are natural and zero-shot speech and singing synthesizers.arXiv preprint arXiv:2304.09116 (2023)

    2023 arXiv

  30. [38]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. InInternational conference on machine learnin...

    2021

  31. [39]

    James Traer, Maddie Cusimano, and Josh H McDermott. 2019. A perceptually inspired generative model of rigid-body contact sounds. InDigital Audio Effects (DAFx), Vol. 1. 3

    2019

  32. [40]

    Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 2024. 4D Gaussian Splatting for Real- Time Dynamic Scene Rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 20310–20320

    2024

  33. [41]

    Kun Su, Kaizhi Qian, Eli Shlizerman, Antonio Torralba, and Chuang Gan. 2023. Physics-driven diffusion models for impact sound synthesis from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 9749–9759

    2023

  34. [42]

    Tengyu Xu, Yingbin Liang, and Guanghui Lan. 2021. Crpo: A new approach for safe reinforcement learning with convergence guarantee. InInternational Conference on Machine Learning. PMLR, 11480–11491

    2021

  35. [43]

    Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li

  36. [44]

    Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. 2024. Point transformer v3: Simpler MM ’25, October 27–31, 2025, Dublin, Ireland. Chunshi Wang, Hongxing Li, & Yawei Luo faster stronger. InProceedings of the I...

    2024

  37. [45]

    Yixiao Zhang, Yukara Ikemiya, Woosung Choi, Naoki Murata, Marco A Martínez- Ramírez, Liwei Lin, Gus Xia, Wei-Hsiang Liao, Yuki Mitsufuji, and Simon Dixon

  38. [46]

    Yipin Zhou, Zhaowen Wang, Chen Fang, Trung Bui, and Tamara L Berg. 2018. Visual to sound: Generating natural sound for videos in the wild. InProceedings of the IEEE conference on computer vision and pattern recognition. 3550–3558

    2018

  39. [47]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Gs-slam: Dense visual slam with 3d gaussian splatting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 19595–19604

  40. [48]

    Zih-Syuan Yang and Jason Hockman. 2023. A plugin for neural audio synthesis of impact sound effects. InProceedings of the 18th International Audio Mostly Conference. 143–146

    2023

  41. [50]

    Instruct-MusicGen: Unlocking Text-to-Music Editing for Music Language Models via Instruction Tuning.arXiv preprint arXiv:2405.18386(2024)

    2024 arXiv

  42. [2023]

    https://repo-sam.inria.fr/fungraph/3d- gaussian-splatting/

    3D Gaussian Splatting for Real-Time Radiance Field Rendering.ACM Transactions on Graphics42, 4 (July 2023). https://repo-sam.inria.fr/fungraph/3d- gaussian-splatting/

    2023

  43. [2024]

    In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Interactdiffusion: Interaction control in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 6180–6189

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.