pith. sign in

arxiv: 2605.29583 · v1 · pith:5EIV24TQnew · submitted 2026-05-28 · 💻 cs.CV

BitC-3DGS: High-Capacity 3D Gaussian Splatting Watermarking via Bit Compression

Yuquan Bi , Baosheng Yu , Yingke Lei , Jianwei Yang , Hongsong Wang , Jie Gui , Yuan Yan Tang , James Tin-Yau Kwok This is my paper

Pith reviewed 2026-06-29 08:21 UTC · model grok-4.3

classification 💻 cs.CV
keywords 3D Gaussian Splattingwatermarkingbit compressionhigh-capacity embeddingmessage recovery3D asset protection
0
0 comments X

The pith

BitC-3DGS compresses multiple bits per token to embed 128-bit watermarks in 3D Gaussian Splatting.

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

The paper introduces BitC-3DGS to overcome the 77-bit limit of existing CLIP-based watermarking for 3D Gaussian Splatting assets. It packs several message bits into each semantic token via a bit-compressed tokenization scheme. A dual-branch architecture then decompresses the chunks and decodes the bits, aided by hard-message sampling during training. Experiments on the Blender and LLFF datasets show that 128-bit messages can be recovered with accuracy comparable to prior 64-bit methods while preserving rendering quality.

Core claim

BitC-3DGS employs a bit-compressed tokenization scheme that encodes multiple bits within the same chunk into a single semantic token, together with a dual-branch architecture for joint chunk decompression and bit decoding and a hard-message sampling strategy, to support 128-bit message capacity with recovery accuracy comparable to that of 64-bit messages in recent state-of-the-art methods.

What carries the argument

Bit-compressed tokenization scheme that encodes multiple bits within the same chunk into a single semantic token, combined with a dual-branch architecture for joint chunk decompression and bit decoding.

If this is right

  • 128-bit messages can be embedded and recovered at accuracy levels matching prior 64-bit work.
  • Rendering fidelity remains high on the Blender and LLFF datasets.
  • Rich information such as ownership, provenance, and authentication codes can be carried inside 3DGS assets.
  • The approach enables reliable identification and integrity checks in large-scale 3D asset pipelines.

Where Pith is reading between the lines

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

  • The same compression idea could be tested on other neural rendering representations that rely on fixed-length text encoders.
  • Further increases in capacity might be possible by varying the number of bits packed per token according to message content.
  • Integration into asset-management pipelines would allow automated provenance checks at render time.

Load-bearing premise

Multiple bits can be packed into one token and later recovered accurately by the dual-branch decoder without substantially harming the rendered image quality.

What would settle it

Embed 128-bit messages into 3DGS scenes on Blender or LLFF and check whether bit recovery accuracy falls below the level reported for 64-bit messages in prior methods or whether PSNR and SSIM degrade noticeably.

Figures

Figures reproduced from arXiv: 2605.29583 by Baosheng Yu, Hongsong Wang, James Tin-Yau Kwok, Jianwei Yang, Jie Gui, Yingke Lei, Yuan Yan Tang, Yuquan Bi.

Figure 1
Figure 1. Figure 1: Decoding accuracy versus message capacity. Conventional image￾space watermarking methods operating on low-level signals (e.g., Gaussian￾Marker [5] and 3D-GSW [6]) suffer notable performance degradation as message capacity increases. Semantic watermarking based on bit-to-token encoding (e.g., GuardSplat [7]) improves decoding accuracy but is inherently limited to 77-bit messages. In contrast, the proposed B… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed BitC-3DGS framework. The method contains two stages. Stage I (decoder pre-training): Messages sampled from the buffer D are converted into compact token sequences via bit-compressed tokenization, encoded by the pre-trained CLIP text encoder, and used to train the dual￾branch decoder DM. During this stage, a hard-message sampling strategy continually updates D by retaining historica… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the bit branch in the proposed dual-branch decoder. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visual comparisons with baselines [5]–[7] under [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative results of BitC-3DGS under 96-bit and 128-bit payloads [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative results of BitC-3DGS under 96-bit and 128-bit payloads [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visual ablation study of the three proposed components on the chair scene (L = 64 bits) from the Blender dataset. Each column shows the rendered result and a magnified patch for a specific configuration. 90.78% to 93.76% while leaving [In] nearly unchanged and reducing the gap from 4.29 to 1.23 points. With all components combined, BitC-3DGS achieves the best balance, reaching 97.90% ([In]) and 96.80% ([Ou… view at source ↗
Figure 8
Figure 8. Figure 8: Sensitivity to hard-message sampling ratios [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
read the original abstract

High-capacity watermarking is necessary for 3D Gaussian Splatting (3DGS) assets to embed rich information (e.g., ownership, provenance, and authentication codes), enabling reliable identification and integrity verification in large-scale 3D asset pipelines. Existing bit-to-token watermarking methods based on a pre-trained text encoder are limited to 77-bit messages due to CLIP's fixed 77-token context length, as tokens beyond this limit are unsupported by learned positional embeddings. To address this limitation, we introduce BitC-3DGS, a bit-compression framework that encodes multiple message bits per token. It employs a bit-compressed tokenization scheme that encodes multiple bits within the same chunk into a single semantic token. To enable recovery of the compressed information, it further introduces a dual-branch architecture for joint chunk decompression and bit decoding, along with a hard-message sampling strategy to improve combinatorial coverage during decoder training. Extensive experiments on the Blender and LLFF datasets demonstrate the effectiveness of BitC-3DGS for high-capacity watermarking, achieving high message recovery accuracy and rendering fidelity. For example, it supports 128-bit message capacity with recovery accuracy comparable to that of 64-bit messages in recent state-of-the-art methods.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces BitC-3DGS, a bit-compression framework for high-capacity watermarking of 3D Gaussian Splatting (3DGS) assets. It overcomes the 77-token limit of pre-trained encoders such as CLIP by proposing a bit-compressed tokenization scheme that packs multiple message bits into each semantic token, paired with a dual-branch decoder for joint chunk decompression and bit recovery, and trained via a hard-message sampling strategy to increase combinatorial coverage. Experiments on the Blender and LLFF datasets are reported to support 128-bit message capacity with recovery accuracy comparable to recent 64-bit state-of-the-art methods while preserving rendering fidelity.

Significance. If the central claims hold, the work enables embedding of richer metadata (ownership, provenance, authentication) into 3DGS assets for large-scale pipelines. The bit-compression approach and dual-branch architecture directly address an external capacity bottleneck of existing encoder-based methods. The hard-message sampling strategy is a concrete engineering contribution to decoder training, though its generalization properties remain to be verified.

major comments (2)
  1. [Abstract] Abstract: the claim that 128-bit messages achieve recovery accuracy comparable to 64-bit SOTA methods rests on the dual-branch decoder generalizing beyond the hard-message samples; the manuscript provides no quantification of the fraction of the 2^128 space covered by sampling nor any evaluation on fully random or adversarial 128-bit messages, which is load-bearing for the reliability assertion.
  2. [Experiments] Experiments section: the abstract states that Blender and LLFF experiments show high recovery accuracy and fidelity, yet supplies no error bars, explicit baselines, data splits, number of trials, or trade-off curves between capacity and PSNR/SSIM; this absence prevents verification of the central performance claim.
minor comments (1)
  1. [Abstract] Abstract: the exact numerical recovery accuracies (e.g., bit-error rate or accuracy percentage) for the 128-bit case versus the 64-bit SOTA references should be stated for immediate comparison.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. We address each major comment below and will revise the manuscript accordingly to improve clarity and rigor.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that 128-bit messages achieve recovery accuracy comparable to 64-bit SOTA methods rests on the dual-branch decoder generalizing beyond the hard-message samples; the manuscript provides no quantification of the fraction of the 2^128 space covered by sampling nor any evaluation on fully random or adversarial 128-bit messages, which is load-bearing for the reliability assertion.

    Authors: We agree that explicit quantification of combinatorial coverage and evaluation on random/adversarial messages would strengthen the generalization claim. The hard-message sampling strategy was designed to prioritize difficult bit combinations during training, but the manuscript does not report coverage fractions or additional random-message tests. In revision we will add (i) an estimate of effective coverage derived from the sampling procedure and (ii) recovery results on fully random 128-bit messages drawn from the full space. revision: yes

  2. Referee: [Experiments] Experiments section: the abstract states that Blender and LLFF experiments show high recovery accuracy and fidelity, yet supplies no error bars, explicit baselines, data splits, number of trials, or trade-off curves between capacity and PSNR/SSIM; this absence prevents verification of the central performance claim.

    Authors: The referee correctly identifies missing experimental details. The current experiments section reports aggregate results on Blender and LLFF but omits error bars, explicit data-split descriptions, trial counts, and capacity-vs-fidelity curves. We will revise the section to include these elements: standard deviations over multiple random seeds, precise train/test splits, number of trials, and trade-off plots relating message capacity to both recovery accuracy and rendering metrics. revision: yes

Circularity Check

0 steps flagged

No circularity identified; claims rest on novel architecture and empirical results

full rationale

The paper proposes BitC-3DGS as a new bit-compression framework to exceed CLIP's 77-token limit via multi-bit tokenization, a dual-branch decoder, and hard-message sampling during training. The 128-bit capacity result is presented as an experimental outcome on Blender and LLFF datasets, with accuracy compared to prior 64-bit methods. No equations, definitions, or claims reduce by construction to fitted inputs or self-citations; the central components are introduced as independent design choices whose effectiveness is externally validated rather than presupposed.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

Only the abstract is available; full details on parameters, assumptions, and evidence are inaccessible.

axioms (1)
  • domain assumption Existing bit-to-token watermarking methods based on a pre-trained text encoder are limited to 77-bit messages due to CLIP's fixed 77-token context length
    Stated directly as the core limitation the work addresses.
invented entities (2)
  • bit-compressed tokenization scheme no independent evidence
    purpose: encodes multiple message bits per token by compressing bits within the same chunk into a single semantic token
    New scheme introduced to bypass token limit
  • dual-branch architecture for joint chunk decompression and bit decoding no independent evidence
    purpose: enables recovery of the compressed information
    New architecture proposed for decoding

pith-pipeline@v0.9.1-grok · 5782 in / 1381 out tokens · 26269 ms · 2026-06-29T08:21:10.020479+00:00 · methodology

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

56 extracted references · 1 canonical work pages · 1 internal anchor

  1. [1]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

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

    2021

  2. [2]

    3d gaussian splatting for real-time radiance field rendering,

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering,”ACM Trans. Graph., vol. 42(4), pp. 1–14, 2023

    2023

  3. [3]

    Mantlemark: Migrating watermarks from multi-view images to radiance fields via frequency modulation,

    Z. Luo, J. Liu, H. Li, A. Rocha, and R. Wan, “Mantlemark: Migrating watermarks from multi-view images to radiance fields via frequency modulation,”IEEE Trans. Inf. F orensics Security, vol. 21, pp. 256–271, 2026

    2026

  4. [4]

    Robust texture-aware local adaptive image watermarking with perceptual guar- antee,

    Y . Huang, H. Guan, J. Liu, S. Zhang, B. Liu, and G. Zhang, “Robust texture-aware local adaptive image watermarking with perceptual guar- antee,”IEEE Trans. Circuit Syst. Video Technol., vol. 33, no. 9, pp. 4660–4674, 2023

    2023

  5. [5]

    Gaussianmarker: Uncertainty-aware copyright protection of 3d gaus- sian splatting,

    X. Huang, R. Li, Y . ming Cheung, K. C. Cheung, S. See, and R. Wan, “Gaussianmarker: Uncertainty-aware copyright protection of 3d gaus- sian splatting,” inNeural Information Processing Systems, 2024, p. 33037–33060

    2024

  6. [6]

    3d-gsw: 3d gaussian splatting watermark for protecting copyrights in radiance fields,

    Y . Jang, H. Park, F. Yang, H. Ko, E. Choo, and S. Kim, “3d-gsw: 3d gaussian splatting watermark for protecting copyrights in radiance fields,” inConference on Computer Vision and Pattern Recognition, 2025, pp. 5938–5948

    2025

  7. [7]

    Guardsplat: Efficient and robust watermarking for 3d gaussian splatting,

    Z. Chen, G. Wang, J. Zhu, J. Lai, and X. Xie, “Guardsplat: Efficient and robust watermarking for 3d gaussian splatting,” inConference on Computer Vision and Pattern Recognition, 2025, pp. 16 325–16 335

    2025

  8. [8]

    A robust reversible watermarking scheme using attack-simulation-based adaptive normaliza- tion and embedding,

    Y . Tang, C. Wang, S. Xiang, and Y .-M. Cheung, “A robust reversible watermarking scheme using attack-simulation-based adaptive normaliza- tion and embedding,”IEEE Trans. Inf. F orensics Security, vol. 19, pp. 4114–4129, 2024

    2024

  9. [9]

    Dual defense: Adversarial, traceable, and invisible robust watermarking against face swapping,

    Y . Zhang, D. Ye, C. Xie, L. Tang, C. Chen, Z. Liu, and J. Deng, “Dual defense: Adversarial, traceable, and invisible robust watermarking against face swapping,”IEEE Trans. Inf. F orensics Security, vol. 19, pp. 4628–4641, 2024

    2024

  10. [10]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, and J. Clark, “Learning transferable visual models from natural language supervision,” inInternational Conference on Learning Representations, 2021, pp. 8748–8763

    2021

  11. [11]

    Improved wavelet-based water- marking through pixel-wise masking,

    M. Barni, F. Bartolini, and A. Piva, “Improved wavelet-based water- marking through pixel-wise masking,”IEEE Trans. Image Process., vol. 10(5), pp. 783–791, 2001

    2001

  12. [12]

    A dwt based watermarking approach for medical image protection,

    F. Kahlessenane, A. Khaldi, R. Kafi, and S. Euschi, “A dwt based watermarking approach for medical image protection,”Journal of Ambi- ent Intelligence and Humanized Computing, vol. 12(2), pp. 2931–2938, 2021

    2021

  13. [13]

    A new dct-based robust im- age watermarking method using teaching-learning-based optimization,

    M. Moosazadeh and G. Ekbatanifard, “A new dct-based robust im- age watermarking method using teaching-learning-based optimization,” Journal of Information Security and Applications, vol. 47, pp. 28–38, 2019

    2019

  14. [14]

    Dwt-dct-svd based watermarking,

    K. Navas, M. C. Ajay, M. Lekshmi, T. S. Archana, and M. Sasiku- mar, “Dwt-dct-svd based watermarking,” inIn 2008 3rd international conference on communication systems software and middleware and workshops, 2008, pp. 271–274

    2008

  15. [15]

    Perceptual dft watermarking with improved detection and robustness to geometrical distortions,

    M. U. D. Goudia and F. Autrusseau, “Perceptual dft watermarking with improved detection and robustness to geometrical distortions,”IEEE Trans. Inf. F orensics Security, vol. 9(7), pp. 1108–1119, 2014

    2014

  16. [16]

    Digital image watermarking using discrete wavelet transform and singular value decomposition,

    C.-C. Lai and C.-C. Tsai, “Digital image watermarking using discrete wavelet transform and singular value decomposition,”IEEE Trans. Instrum. Meas., vol. 59(11), pp. 3060–3063, 2010

    2010

  17. [17]

    Hidden: Hiding data with deep networks,

    J. Zhu, R. Kaplan, J. Johnson, and L. Fei-Fei, “Hidden: Hiding data with deep networks,” inEuropean Conference on Computer Vision, 2018, pp. 682–697

    2018

  18. [18]

    Towards blind watermarking: Combining invertible and non-invertible mecha- nisms,

    R. Ma, M. Guo, Y . Hou, F. Yang, Y . Li, H. Jia, and X. Xie, “Towards blind watermarking: Combining invertible and non-invertible mecha- nisms,” inACM International Conference on Multimedia, 2022, pp. 1532–1542

    2022

  19. [19]

    Stegastamp: Invisible hyperlinks in physical photographs,

    M. Tancik, B. Mildenhall, and R. Ng, “Stegastamp: Invisible hyperlinks in physical photographs,” inConference on Computer Vision and Pattern Recognition, 2020, pp. 2117–2126

    2020

  20. [20]

    Red- mark: Framework for residual diffusion watermarking based on deep networks,

    M. Ahmadi, A. Norouzi, N. Karimi, S. Samavi, and A. Emami, “Red- mark: Framework for residual diffusion watermarking based on deep networks,”Expert Systems with Applications, vol. 146, p. 113157, 2020

    2020

  21. [21]

    High-capacity convolutional video steganography with temporal residual modeling,

    X. Weng, Y . Li, L. Chi, and Y . Mu, “High-capacity convolutional video steganography with temporal residual modeling,” inICMR, 2019, pp. 87–95

    2019

  22. [22]

    Joint watermarking-encryption-jpeg-ls for medical image reliability control in encrypted and compressed domains,

    S. Haddad, G. Coatrieux, A. Moreau-Gaudry, and M. Cozic, “Joint watermarking-encryption-jpeg-ls for medical image reliability control in encrypted and compressed domains,”IEEE Trans. Inf. F orensics Security, vol. 15, pp. 2556–2569, 2020

    2020

  23. [23]

    Imprints: Mitigating watermark removal attacks with defensive watermarks,

    X. Chen, J. Deng, Y . Chen, C. Li, X. Fang, C. Liu, and W. Xu, “Imprints: Mitigating watermark removal attacks with defensive watermarks,”IEEE Trans. Inf. F orensics Security, vol. 20, pp. 1866–1881, 2025

    2025

  24. [24]

    Screen-shooting resilient watermarking,

    H. Fang, W. Zhang, H. Zhou, H. Cui, and N. Yu, “Screen-shooting resilient watermarking,”IEEE Trans. Inf. F orensics Security, vol. 14, no. 6, pp. 1403–1418, 2019

    2019

  25. [25]

    Screen- shooting robust watermark based on style transfer and structural re- parameterization,

    G. Gao, X. Chen, L. Li, Z. Xia, J. Fei, and Y .-Q. Shi, “Screen- shooting robust watermark based on style transfer and structural re- parameterization,”IEEE Trans. Inf. F orensics Security, vol. 20, pp. 2648–2663, 2025

    2025

  26. [26]

    Gaus- sian shading: Provable performance-lossless image watermarking for diffusion models,

    Z. Yang, K. Zeng, K. Chen, H. Fang, W. Zhang, and N. Yu, “Gaus- sian shading: Provable performance-lossless image watermarking for diffusion models,” inConference on Computer Vision and Pattern Recognition, 2024, pp. 12 162–12 171

    2024

  27. [27]

    Omniguard: Hybrid manipulation localization via augmented versatile deep image watermarking,

    X. Zhang, Z. Tang, Z. Xu, R. Li, Y . Xu, B. Chen, F. Gao, and J. Zhang, “Omniguard: Hybrid manipulation localization via augmented versatile deep image watermarking,” inConference on Computer Vision and Pattern Recognition, 2025, pp. 3008–3018

    2025

  28. [28]

    The stable signature: Rooting watermarks in latent diffusion models,

    P. Fernandez, G. Couairon, H. J ´egou, M. Douze, and T. Furon, “The stable signature: Rooting watermarks in latent diffusion models,” in International Conference on Computer Vision, 2023, pp. 22 466–22 477

    2023

  29. [29]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” inNeural Information Processing Systems, 2020, pp. 6840–6851

    2020

  30. [30]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” inConference on Computer Vision and Pattern Recognition, 2022, pp. 10 684–10 695

    2022

  31. [31]

    Tensorf: Tensorial radiance fields,

    A. Chen, Z. Xu, A. Geiger, J. Yu, and H. Su, “Tensorf: Tensorial radiance fields,” inEuropean Conference on Computer Vision, 2022, pp. 333–350

    2022

  32. [32]

    Plenoxels: Radiance fields without neural networks,

    S. Fridovich-Keil, A. Yu, M. Tancik, Q. Chen, B. Recht, and A. Kanazawa, “Plenoxels: Radiance fields without neural networks,” inConference on Computer Vision and Pattern Recognition, 2022, pp. 5501–5510

    2022

  33. [33]

    Instant neural graphics primitives with a multiresolution hash encoding,

    T. M ¨uller, A. Evans, C. Schied, and A. Keller, “Instant neural graphics primitives with a multiresolution hash encoding,”ACM Trans. Graph., vol. 41(4), pp. 1–15, 2022

    2022

  34. [34]

    Copyrnerf: Protecting the copyright of neural radiance fields,

    Z. Luo, Q. Guo, K. C. Cheung, S. See, and R. Wan, “Copyrnerf: Protecting the copyright of neural radiance fields,” inInternational Conference on Computer Vision, 2023, pp. 22 401–22 411

    2023

  35. [35]

    Steganerf: Embedding invisible information within neural radiance fields,

    C. Li, B. Y . Feng, Z. Fan, P. Pan, and Z. Wang, “Steganerf: Embedding invisible information within neural radiance fields,” inInternational Conference on Computer Vision, 2023, pp. 441–453

    2023

  36. [36]

    Waterf: Robust watermarks in radiance fields for protection of copyrights,

    Y . Jang, D. I. Lee, M. Jang, J. W. Kim, F. Yang, and S. Kim, “Waterf: Robust watermarks in radiance fields for protection of copyrights,” in Conference on Computer Vision and Pattern Recognition, 2024, pp. 12 087–12 097

    2024

  37. [37]

    Fate: Full-head gaussian avatar with textural editing from monocular video,

    J. Zhang, Z. Wu, Z. Liang, Y . Gong, D. Hu, Y . Yao, X. Cao, and H. Zhu, “Fate: Full-head gaussian avatar with textural editing from monocular video,” inConference on Computer Vision and Pattern Recognition, 2025, pp. 5535–5545

    2025

  38. [38]

    Anigs: Animatable gaussian avatar from a single image with inconsistent gaussian reconstruction,

    L. Qiu, S. Zhu, Q. Zuo, X. Gu, Y . Dong, J. Zhang, C. Xu, Z. Li, W. Yuan, L. Bo, G. Chen, and Z. Dong, “Anigs: Animatable gaussian avatar from a single image with inconsistent gaussian reconstruction,” inConference on Computer Vision and Pattern Recognition, 2025, pp. 21 148–21 158

    2025

  39. [39]

    Real-time high-fidelity gaussian human avatars with position-based interpolation of spatially distributed mlps,

    Y . Zhan, T. Shao, Y . Yang, and K. Zhou, “Real-time high-fidelity gaussian human avatars with position-based interpolation of spatially distributed mlps,” inConference on Computer Vision and Pattern Recognition, 2025, pp. 26 297–26 307

    2025

  40. [40]

    Moga: 3d generative avatar prior for monocular gaussian avatar reconstruction,

    Z. Dong, L. Duan, J. Song, M. J. Black, and A. Geiger, “Moga: 3d generative avatar prior for monocular gaussian avatar reconstruction,” in International Conference on Computer Vision, 2025, pp. 13 304–13 314

    2025

  41. [41]

    Objectgs: Object-aware scene reconstruction and scene understanding via gaussian splatting,

    R. Zhu, M. Yu, L. Xu, L. Jiang, Y . Li, T. Zhang, J. Pang, and B. Dai, “Objectgs: Object-aware scene reconstruction and scene understanding via gaussian splatting,” inInternational Conference on Computer Vision, 2025, pp. 8350–8360

    2025

  42. [42]

    Robust and efficient 3d gaussian splatting for urban scene reconstruction,

    Z. Yuan, H. Huang, Z. Xiong, D. Wang, and G. Yang, “Robust and efficient 3d gaussian splatting for urban scene reconstruction,” in International Conference on Computer Vision, 2025, pp. 26 209–26 219. MANUSCRIPT FOR IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY 12

    2025

  43. [43]

    Physsplat: Efficient physics simulation for 3d scenes via mllm-guided gaussian splatting,

    H. Zhao, H. Wang, X. Zhao, H. Fei, H. Wang, C. Long, and H. Zou, “Physsplat: Efficient physics simulation for 3d scenes via mllm-guided gaussian splatting,” inInternational Conference on Computer Vision, 2025, pp. 5242–5252

    2025

  44. [44]

    Motiongs: Exploring explicit motion guidance for deformable 3d gaussian splatting,

    R. Zhu, Y . Liang, H. Chang, J. Deng, J. Lu, W. Yang, T. Zhang, and Y . Zhang, “Motiongs: Exploring explicit motion guidance for deformable 3d gaussian splatting,” inNeural Information Processing Systems, 2025, p. 101790–101817

    2025

  45. [45]

    Generative sparse-view gaussian splatting,

    H. Kong, X. Yang, and X. Wang, “Generative sparse-view gaussian splatting,” inConference on Computer Vision and Pattern Recognition, 2025, pp. 26 745–26 755

    2025

  46. [46]

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

    T. Yi, J. Fang, J. Wang, G. Wu, L. Xie, X. Zhang, W. Liu, Q. Tian, and X. Wang, “Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models,” inConference on Computer Vision and Pattern Recognition, 2024, pp. 6796–6807

    2024

  47. [47]

    Uvgs: Reimagining unstructured 3d gaussian splatting using uv mapping,

    A. Rai, D. Wang, M. Jain, N. Sarafianos, K. Chen, S. Sridhar, and A. Prakash, “Uvgs: Reimagining unstructured 3d gaussian splatting using uv mapping,” inConference on Computer Vision and Pattern Recognition, 2025, pp. 5927–5937

    2025

  48. [48]

    Directtrigs: Triplane-based gaussian splatting field representation for 3d generation,

    X. Ju and H. Li, “Directtrigs: Triplane-based gaussian splatting field representation for 3d generation,” inConference on Computer Vision and Pattern Recognition, 2025, pp. 16 229–16 239

    2025

  49. [49]

    Gs-hider: Hiding messages into 3d gaussian splatting,

    X. Zhang, J. Meng, R. Li, Z. Xu, Y . Zhang, and J. Zhang, “Gs-hider: Hiding messages into 3d gaussian splatting,” inNeural Information Processing Systems, 2024, p. 49780–49805

    2024

  50. [50]

    Instantsplamp: Fast and generalizable stenography framework for generative gaussian splatting,

    C. Li, H. Liu, Z. Fan, W. Li, Y . Liu, P. Pan, and Y . Yuan, “Instantsplamp: Fast and generalizable stenography framework for generative gaussian splatting,” inInternational Conference on Learning Representations, 2025

    2025

  51. [51]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Conference on Computer Vision and Pattern Recognition, 2018, pp. 586– 595

    2018

  52. [52]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszko-reit, L. Jones, A. N. Gomez, Łukasz Kaiser, and I. Polosukhin, “Attention is all you need,” inNeural Information Processing Systems, 2017, pp. 5998–6008

    2017

  53. [53]

    Local light field fusion: Practical view synthesis with prescriptive sampling guidelines,

    B. Mildenhall, P. P. Srinivasan, R. Ortiz-Cayon, N. K. Kalantari, R. Ra- mamoorthi, R. Ng, and A. Kar, “Local light field fusion: Practical view synthesis with prescriptive sampling guidelines,”ACM Trans. Graph., vol. 38(4), pp. 1–14, 2019

    2019

  54. [54]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,”IEEE Trans. Image Process., vol. 13(4), pp. 600–612, 2004

    2004

  55. [55]

    Adam: A Method for Stochastic Optimization

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  56. [56]

    Differentiable jpeg: The devil is in the details,

    C. Reich, B. Debnath, D. Patel, and S. Chakradhar, “Differentiable jpeg: The devil is in the details,” inWinter Conference on Applications of Computer Vision, 2024, pp. 4126–4135

    2024