pith. sign in

arxiv: 2606.24144 · v1 · pith:ITB7XLP2new · submitted 2026-06-23 · 💻 cs.CV

Geometry-Aware Style Transfer in 3D Gaussian Splatting

Min Hyeok Bang , Jun Hyeong Kim , Seung-Wook Kim , Se-Ho Lee This is my paper

Pith reviewed 2026-06-26 01:40 UTC · model grok-4.3

classification 💻 cs.CV
keywords style transfer3D Gaussian splattinggeometry adaptationcontrastive feature matchingdecoupled optimizationscene stylizationstructural transfer
0
0 comments X

The pith

A decoupled optimization scheme alternately updates color and geometry to transfer both appearance and structure in 3D Gaussian splatting.

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

This paper introduces a style transfer method for 3D scenes modeled with Gaussian splatting that changes both the colors and the underlying shapes. It separates the updates so that color changes and geometry changes happen in turn rather than together. The separation relies on a matching process that compares features from color, depth, and edge maps between the current scene and the style image. This leads to more reliable changes in the 3D structure without the updates fighting each other. A reader would care because it opens the door to style transfers that actually reshape objects instead of just recoloring them.

Core claim

Our method explicitly incorporates geometry adaptation through a decoupled optimization scheme that alternately updates color and geometry parameters. This strategy alleviates potential interference between color and geometry updates, leading to stable and consistent scene-level geometry transformation. The decoupled optimization is enabled by the proposed geometry-aware contrastive feature matching (GCFM). GCFM integrates RGB, depth, and edge cues into a contrastive objective and is employed in both optimization phases to effectively transfer structural characteristics from style images to Gaussian primitives.

What carries the argument

Decoupled optimization scheme that alternately updates color and geometry parameters, enabled by geometry-aware contrastive feature matching (GCFM) that integrates RGB, depth, and edge cues into a contrastive objective.

If this is right

  • Stable and consistent scene-level geometry transformation occurs without interference between color and geometry updates.
  • Superior performance is achieved in both qualitative fidelity and quantitative metrics compared to prior methods.
  • Simultaneous transfer of appearance attributes and geometric structures becomes possible in 3DGS scenes.
  • Existing 3DGS-based stylization methods are significantly outperformed on structural adaptation tasks.

Where Pith is reading between the lines

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

  • The method could extend to dynamic or video-based 3D scenes if temporal consistency is added to the decoupled updates.
  • Structural style transfer might improve applications like virtual object redesign where shape changes matter more than recoloring.
  • Similar contrastive matching on multiple cues could apply to other 3D representations if the Gaussian primitive assumption is relaxed.

Load-bearing premise

The geometry-aware contrastive feature matching successfully transfers structural characteristics from style images to Gaussian primitives without introducing inconsistencies or artifacts.

What would settle it

A test scene where the style image has substantially different depth and edge structures produces visible geometric artifacts or inconsistencies after optimization.

Figures

Figures reproduced from arXiv: 2606.24144 by Jun Hyeong Kim, Min Hyeok Bang, Se-Ho Lee, Seung-Wook Kim.

Figure 1
Figure 1. Figure 1: Examples of geometry-aware style transfer in 3D Gaussian splatting (3DGS). Given an input scene and various style guides, our method stylizes both color and geometry while preserving overall scene structure. Highlighted depth regions show how geometric details adaptively reflect structural cues from the target style. results on single-view images, they are inherently limited to 2D representations and canno… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed geometry-aware style transfer framework for 3DGS. The pipeline consists of (a) 3DGS representation, (b) color matching, and (c) decoupled optimization between color and geometry optimization phases guided by multi-modal (color, depth, edge) features from (d) the style reference via the proposed GCFM [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison between joint-step (top) and decoupled (bottom) optimization. The joint-step optimization updates color and geometry simultaneously, while our de￾coupled optimization alternates them (both run for the same total iterations). 3DGS representation Ginit = {Gn := (θn, µn, Σn, αn)} N n=1 is reconstructed from I C , where each Gaussian primitive Gn is defined by the color representation θn, its 3D cen… view at source ↗
Figure 4
Figure 4. Figure 4: Conceptual illustration of the GCFM process. for Ng inner steps. For j = 0, . . . , Ng − 1, the geometry parameters are updated as \mathbf {\Phi }^{k,j+1} = \mathbf {\Phi }^{k,j} - \eta \nabla _{\mathbf {\Phi }} \mathcal {L}(\mathbf {\Theta }^{k+1}, \mathbf {\Phi }^{k,j}). \label {eq:geo_update} (4) The resulting geometry parameters, Φk+1 = Φk,Ng are then carried forward to the next outer cycle, completing… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of the Flower and Horns scenes. The top row shows the input scenes, rendered depth maps, and the corresponding style image for each scene. The subsequent rows present stylized results produced by different methods, including G-Style [18], StylizedGS [36], SGSST [4], CLIPGaussian [8], and Ours. 4 Experiments 4.1 Datasets We evaluate our 3DGS-based style transfer method across a total … view at source ↗
Figure 6
Figure 6. Figure 6: Results of the user study comparing StyleGaussian [21], G-Style [18], SGSST [4], CLIPGaussian [8], and Ours. Quantitative comparison [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Ablation study on GCFM. The 1st and 3rd columns show the stylized rendered views, while the 2nd and 4th columns display the corresponding enlarged depth maps. demonstrating its effectiveness in producing high-quality stylizations that are faithful to the target style while preserving scene content. 4.5 Complexity Analysis [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
read the original abstract

In this paper, we present a novel geometry-aware style transfer framework for 3D Gaussian splatting (3DGS) that simultaneously transfers appearance attributes and geometric structures. Unlike prior works that primarily focus on color-based stylization and often overlook structural adaptation, our method explicitly incorporates geometry adaptation through a decoupled optimization scheme that alternately updates color and geometry parameters. This strategy alleviates potential interference between color and geometry updates, leading to stable and consistent scene-level geometry transformation. The decoupled optimization is enabled by the proposed geometry-aware contrastive feature matching (GCFM). GCFM integrates RGB, depth, and edge cues into a contrastive objective and is employed in both optimization phases to effectively transfer structural characteristics from style images to Gaussian primitives. Extensive experiments show that our approach achieves superior performance in both qualitative fidelity and quantitative metrics, significantly outperforming existing 3DGS-based stylization methods. Our code is available at \href{https://github.com/oweixx/gast}{https://github.com/oweixx/gast}.

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 / 2 minor

Summary. The paper presents a geometry-aware style transfer framework for 3D Gaussian Splatting that transfers both appearance and geometric structure from style images. It introduces a decoupled alternating optimization scheme for color and geometry parameters, enabled by a geometry-aware contrastive feature matching (GCFM) objective that integrates RGB, depth, and edge cues to guide updates on Gaussian primitives. The method is claimed to reduce interference between color and geometry, yielding stable scene-level transformations, with experiments asserting superior qualitative and quantitative results over prior 3DGS stylization approaches. Code is released at the provided GitHub link.

Significance. If the empirical claims hold, the work addresses an under-explored aspect of 3D stylization by explicitly handling geometry adaptation rather than color-only transfer. The decoupled optimization and multi-cue contrastive matching represent a practical algorithmic pattern that could generalize to other 3D representation tasks. The public code release supports reproducibility and further investigation.

major comments (2)
  1. [Experiments] Experiments section: the central claim of superior performance requires explicit reporting of quantitative metrics (e.g., PSNR, LPIPS, or geometry-specific measures), baselines, and ablation tables; without these details the assertion that the decoupled scheme and GCFM produce measurable gains cannot be evaluated.
  2. [Method] §3.2 (GCFM formulation): the contrastive objective combining RGB, depth, and edge cues is described at a high level but lacks the precise loss equation or weighting scheme; this is load-bearing for verifying that the matching transfers structure without introducing inconsistencies.
minor comments (2)
  1. [Abstract] The abstract and introduction could more clearly distinguish the proposed GCFM from standard contrastive losses used in prior style transfer works.
  2. [Figures] Figure captions should explicitly state which scenes and style images are shown to allow direct comparison with the quantitative claims.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below and will revise the manuscript accordingly to improve clarity and completeness.

read point-by-point responses

  1. Referee: [Experiments] Experiments section: the central claim of superior performance requires explicit reporting of quantitative metrics (e.g., PSNR, LPIPS, or geometry-specific measures), baselines, and ablation tables; without these details the assertion that the decoupled scheme and GCFM produce measurable gains cannot be evaluated.

    Authors: We agree that explicit quantitative reporting is necessary to substantiate the claims. The current manuscript mentions quantitative metrics in the abstract and experiments but does not present them in dedicated tables with all baselines and ablations. In the revision we will add comprehensive tables reporting PSNR, LPIPS, depth error, and other geometry measures, full baseline comparisons, and ablation studies isolating the decoupled optimization and GCFM contributions. revision: yes

  2. Referee: [Method] §3.2 (GCFM formulation): the contrastive objective combining RGB, depth, and edge cues is described at a high level but lacks the precise loss equation or weighting scheme; this is load-bearing for verifying that the matching transfers structure without introducing inconsistencies.

    Authors: We acknowledge the need for the exact formulation. Section 3.2 currently describes GCFM at a high level. We will insert the full mathematical definition of the contrastive loss, the precise combination of RGB, depth, and edge terms, and the weighting coefficients used in the revised manuscript. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper describes an algorithmic framework consisting of a decoupled alternating optimization scheme (color then geometry parameters) enabled by a geometry-aware contrastive feature matching loss (GCFM) that fuses RGB, depth, and edge cues. No equations, derivations, or predictions are shown that reduce any claimed output to a fitted input or self-referential definition by construction. The central claim is presented as an independent engineering contribution whose validity rests on external empirical results rather than internal reduction; no self-citation chains or ansatzes imported from prior author work are invoked as load-bearing steps in the provided text.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the untested effectiveness of the newly introduced GCFM component and the assumption that alternating updates will remain stable across diverse scenes and style images.

axioms (1)
  • domain assumption Style images contain transferable geometric structure that can be captured by RGB, depth, and edge cues and matched to Gaussian primitives via contrastive loss.
    Invoked when the abstract states that GCFM is employed to transfer structural characteristics from style images.
invented entities (1)
  • Geometry-aware contrastive feature matching (GCFM) no independent evidence
    purpose: Integrates RGB, depth, and edge cues into a contrastive objective to guide geometry transfer during optimization.
    New component introduced to enable the decoupled geometry adaptation; no independent evidence outside the paper is provided.

pith-pipeline@v0.9.1-grok · 5717 in / 1356 out tokens · 24546 ms · 2026-06-26T01:40:26.992200+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

40 extracted references · 1 canonical work pages

  1. [1]

    In: CVPR

    Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In: CVPR. pp. 5470–5479 (2022)

    2022

  2. [2]

    IEEE TPAMIP AMI- 8(6), 679–698 (1986)

    Canny, J.: A computational approach to edge detection. IEEE TPAMIP AMI- 8(6), 679–698 (1986)

    1986

  3. [3]

    com / premium - vector / seamless - pattern - with - water - waves - splashes _ 9413868.htm

    Freepik: Seamless pattern with water waves and splashes.https://www.freepik. com / premium - vector / seamless - pattern - with - water - waves - splashes _ 9413868.htm

  4. [4]

    In: CVPR

    Galerne, B., Wang, J., Raad, L., Morel, J.M.: SGSST: Scaling Gaussian splatting style transfer. In: CVPR. pp. 26535–26544 (June 2025)

    2025

  5. [5]

    In: CVPR

    Gatys, L.A., Ecker, A.S., Bethge, M.: Image style transfer using convolutional neural networks. In: CVPR. pp. 2414–2423 (2016)

    2016

  6. [6]

    In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R

    Heusel,M.,Ramsauer,H.,Unterthiner,T.,Nessler,B.,Hochreiter,S.:Ganstrained by a two time-scale update rule converge to a local nash equilibrium. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) NeurIPS. vol. 30. Curran Associates, Inc. (2017)

    2017

  7. [7]

    In: CVPR

    Höllein, L., Johnson, J., Nießner, M.: StyleMesh: Style transfer for indoor 3D scene reconstructions. In: CVPR. pp. 6198–6208 (2022)

    2022

  8. [8]

    In: NeurIPS (2025)

    Howil,K.,Waczyńska,J.,Borycki,P.,Dziarmaga,T.,Mazur,M.,Spurek,P.:CLIP- Gaussian: Universal and multimodal style transfer based on Gaussian splatting. In: NeurIPS (2025)

    2025

  9. [9]

    In: ICCV

    Huang, H.P., Tseng, H.Y., Saini, S., Singh, M., Yang, M.H.: Learning to stylize novel views. In: ICCV. pp. 13869–13878 (October 2021)

    2021

  10. [10]

    In: ICCV

    Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: ICCV. pp. 1501–1510 (2017)

    2017

  11. [11]

    In: ECCV

    Huang, X., Liu, M.Y., Belongie, S., Kautz, J.: Multimodal unsupervised image-to- image translation. In: ECCV. pp. 172–189 (2018)

    2018

  12. [12]

    In: CVPR

    Huang, Y.H., He, Y., Yuan, Y.J., Lai, Y.K., Gao, L.: StylizedNeRF: Consistent 3D scene stylization as stylized NeRF via 2D-3D mutual learning. In: CVPR. pp. 18342–18352 (June 2022)

    2022

  13. [13]

    In: ECCV

    Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: ECCV. pp. 694–711 (2016)

    2016

  14. [14]

    In: CVPR

    Jung, H., Nam, S., Sarafianos, N., Yoo, S., Sorkine-Hornung, A., Ranjan, R.: Ge- ometry transfer for stylizing radiance fields. In: CVPR. pp. 8565–8575 (June 2024)

    2024

  15. [15]

    ACM Transactions on Graphics42(4) (July 2023)

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3D Gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics42(4) (July 2023)

    2023

  16. [16]

    In: Interna- tional Conference on Learning Representations (ICLR) (2015)

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: Interna- tional Conference on Learning Representations (ICLR) (2015)

    2015

  17. [17]

    ACM TOG36(4), 1–13 (2017) 16 M

    Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM TOG36(4), 1–13 (2017) 16 M. H. Bang et al

    2017

  18. [18]

    In: Computer Graphics Forum

    Kovács, Á.S., Hermosilla, P., Raidou, R.G.: G-Style: Stylized Gaussian splatting. In: Computer Graphics Forum. vol. 43, p. e15259. Wiley Online Library (2024)

    2024

  19. [19]

    In: ECCV

    Lee, H.Y., Tseng, H.Y., Huang, J.B., Singh, M., Yang, M.H.: Diverse image-to- image translation via disentangled representations. In: ECCV. pp. 35–51 (2018)

    2018

  20. [20]

    In: CVPR

    Liu, K., Zhan, F., Chen, Y., Zhang, J., Yu, Y., El Saddik, A., Lu, S., Xing, E.P.: StyleRF: Zero-shot 3D style transfer of neural radiance fields. In: CVPR. pp. 8338– 8348 (June 2023)

    2023

  21. [21]

    In: SIGGRAPH Asia (2024)

    Liu, K., Zhan, F., Xu, M., Theobalt, C., Shao, L., Lu, S.: StyleGaussian: Instant 3D style transfer with Gaussian splatting. In: SIGGRAPH Asia (2024)

    2024

  22. [22]

    In: ICME (2025)

    Liu, W., Liu, Z., Yang, X., Sha, M., Li, Y.: ABC-GS: Alignment-based controllable style transfer for 3D Gaussian splatting. In: ICME (2025)

    2025

  23. [23]

    ACM TOG38(4) (2019)

    Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A.: Local light field fusion: Practical view synthesis with prescrip- tive sampling guidelines. ACM TOG38(4) (2019)

    2019

  24. [24]

    In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) ECCV. pp. 405–421 (2020)

    2020

  25. [25]

    ACM Trans

    Nguyen-Phuoc, T., Liu, F., Xiao, L.: SNeRF: Stylized neural implicit representa- tions for 3D scenes. ACM Trans. Graph.41(4) (Jul 2022)

    2022

  26. [26]

    In: CVPR

    Niklaus, S., Liu, F.: Softmax splatting for video frame interpolation. In: CVPR. pp. 5437–5446 (June 2020)

    2020

  27. [27]

    In: NeurIPS (2019)

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: PyTorch: An imperative style, high- performance deep learning library. In: NeurIPS (2019)

    2019

  28. [28]

    In: ICCV

    Shaham, T.R., Dekel, T., Michaeli, T.: SinGAN: Learning a generative model from a single natural image. In: ICCV. pp. 4570–4580 (October 2019)

    2019

  29. [29]

    In: ICLR (2015)

    Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: ICLR (2015)

    2015

  30. [30]

    IEEE Trans

    Tan, W.R., Chan, C.S., Aguirre, H., Tanaka, K.: Improved ArtGAN for conditional synthesis of natural image and artwork. IEEE TIP28(1), 394–409 (2019),https: //doi.org/10.1109/TIP.2018.2866698

    work page doi:10.1109/tip.2018.2866698 2019

  31. [31]

    In: ECCV

    Teed, Z., Deng, J.: RAFT: Recurrent all-pairs field transforms for optical flow. In: ECCV. p. 402–419 (2020)

    2020

  32. [32]

    CoRRabs/1607.08022(2016)

    Ulyanov, D., Vedaldi, A., Lempitsky, V.S.: Instance normalization: The missing ingredient for fast stylization. CoRRabs/1607.08022(2016)

    Pith/arXiv arXiv 2016

  33. [33]

    In: ICCV

    Wang, Z., Zhao, L., Xing, W.: StyleDiffusion: Controllable disentangled style trans- fer via diffusion models. In: ICCV. pp. 7677–7689 (October 2023)

    2023

  34. [34]

    CoRRabs/2406.09414(2024)

    Yang, L., Kang, B., Huang, Z., Zhao, Z., Xu, X., Feng, J., Zhao, H.: Depth anything V2. CoRRabs/2406.09414(2024)

    Pith/arXiv arXiv 2024

  35. [35]

    In: 2024 International Conference on 3D Vision (3DV)

    Zhang, D., Fernandez-Labrador, C., Schroers, C.: CoARF: Controllable 3D artistic style transfer for radiance fields. In: 2024 International Conference on 3D Vision (3DV). pp. 612–622 (2024)

    2024

  36. [36]

    IEEE TPAMI pp

    Zhang, D., Yuan, Y.J., Chen, Z., Zhang, F.L., He, Z., Shan, S., Gao, L.: StylizedGS: Controllable stylization for 3D Gaussian splatting. IEEE TPAMI pp. 1–13 (2025)

    2025

  37. [37]

    In: ECCV

    Zhang, K., Kolkin, N., Bi, S., Luan, F., Xu, Z., Shechtman, E., Snavely, N.: ARF: Artistic radiance fields. In: ECCV. pp. 717–733 (2022)

    2022

  38. [38]

    In: CVPR

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586–595 (2018)

    2018

  39. [39]

    Zhang, Y., He, Z., Xing, J., Yao, X., Jia, J.: Ref-NPR: Reference-based non- photorealisticradiancefieldsforcontrollablescenestylization.In:CVPR.pp.4242– 4251 (June 2023) Geometry-Aware Style Transfer in 3D Gaussian Splatting 17

    2023

  40. [40]

    Please rank the following results according to their similarity to the reference style (1=Most Similar, 5=Least Similar)

    Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: ICCV. pp. 2242–2251 (2017) Geometry-Aware Style Transfer in 3D Gaussian Splatting 1 Geometry-Aware Style Transfer in 3D Gaussian Splatting – Supplementary Material – Algorithm AGeometry-aware style transfer in 3DGS 1:Input:Cont...

    arXiv 2017