We present Stylos, a single-forward 3D Gaussian framework for 3D style transfer that operates on unposed content, from a single image to a multi-view collection, conditioned on a separate reference style image. Stylos synthesizes a stylized 3D Gaussian scene without per-scene optimization or precomputed poses, achieving geometry-aware, view-consistent stylization that generalizes to unseen categories, scenes, and styles. At its core, Stylos adopts a Transformer backbone with two pathways: geometry predictions retain self-attention to preserve geometric fidelity, while style is injected via global cross-attention to enforce visual consistency across views. With the addition of a voxel-based 3D style loss that aligns aggregated scene features to style statistics, Stylos enforces view-consistent stylization while preserving geometry. Experiments across multiple datasets demonstrate that Stylos delivers high-quality zero-shot stylization, highlighting the effectiveness of global style-content coupling, the proposed 3D style loss, and the scalability of our framework from single view to large-scale multi-view settings. Our codes are available at https://github.com/HanzhouLiu/Stylos.
Figures & tables
Figure 1: Architecture overview . Given multi-view content inputs and a style reference, Stylos enables instant 3D stylization without scene-specific training or post-optimization. Its core CrossBlock module facilitates style injection by integrating a cross-attention layer between self-attention and MLP. The proposed 3D style loss matches voxelized 3D features with 2D style statistics.
Strategies
Skateboard
Pizza
Donut
Global
Frame
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
✓
✓
21.12
0.6858
0.2821
19.78
0.5939
0.3326
21.39
0.7198
0.3264
✓
20.93
0.6917
0.2912
19.72
0.5940
0.3405
21.40
0.7167
0.3340
✓
21.68
0.7043
0.2684
20.57
0.6177
0.3110
22.09
0.7362
0.3125
Table 1: Ablation on the style-content fusion module, comparing Frame, Global and hybrid CrossBlock designs. The first frame of each content scene is used as the pseudo style reference. Reconstruction quality is evaluated with PSNR ↑ , SSIM ↑ , and LPIPS ↓ on the CO3D dataset.
Figure 2: CO3D pizza scene comparing different CrossBlock designs.
Figure 3: Comparison of style losses on unseen donut, skateboard, and pizza scenes from the CO3D dataset. Both scene and 3D style losses yield cleaner stylized textures compared to image-level matching, while the 3D loss further conveys a stronger sense of 3D geometry. We encourage readers to Appendix Fig. 8 - 10 for more visual comparisons under varing scenes and styles.
Style Loss
Short-range
Long-range
ArtScore ↑
LPIPS ↓
RMSE ↓
LPIPS ↓
RMSE ↓
Image loss (Eq. 3 )
0.048
0.038
0.157
0.142
4.78
Scene loss (Eq. 4 )
0.047
0.036
0.156
0.148
9.12
3D loss (Eq. 5 )
0.047
0.034
0.153
0.142
9.15
Table 2: Comparison of consistency and artistic quality among different style loss designs on CO3D. The frame stride is set to 3, and 15 held-out scenes are randomly selected for evaluation.
Figure 4: Effect of varying # views / batch on the Lighthouse scene from Tanks and Temples.
Method
Train
Truck
M60
Garden
LPIPS ↓
RMSE ↓
LPIPS ↓
RMSE ↓
LPIPS ↓
RMSE ↓
LPIPS ↓
RMSE ↓
Short-range consistency
StyleGaussian 2024
0.033
0.038
0.031
0.034
0.038
0.037
0.069
0.061
G-Style 2024
0.042
0.052
0.032
0.035
0.038
0.034
0.066
0.070
SGSST 2025
0.038
0.047
0.039
0.047
0.044
0.049
0.084
0.090
Styl3R 2025b
–
–
0.061
0.036
0.066
0.040
0.105
0.067
Table 3: For consistency comparisons, we report short-range and long-range LPIPS ↓ and RMSE ↓ on the four scenes from the Tanks & Temples dataset. We clarify experiment details in A.4 . In the following tables, the best results are highlighted and the second best results are underlined . Each stylization method category is visualized with a distinct color. The proposed Stylos demonstrates improved short-range and long-range consistency scores across the four scenes.
Method
Train
Truck
M60
Garden
Time ↓
Score ↑
FID ↓
Score ↑
FID ↓
Score ↑
FID ↓
Score ↑
FID ↓
StyleGaussian 2024
0.78
52.79
5.76
44.93
8.63
47.48
9.38
41.14
165 m *
G-Style 2024
9.52
23.24
9.67
22.15
9.73
22.36
8.98
25.76
14.7 m *
SGSST 2025
1.84
38.24
5.34
32.34
5.26
38.73
4.89
33.54
35.2 m *
Styl3R 2025b
–
–
2.94
34.11
2.96
29.86
4.09
38.28
0.16 s †
Stylos (ours)
9.50
26.40
9.70
28.71
9.37
27.44
9.34
28.06
0.05 s †
Table 4: Stylization quality, as measured by ArtScore and ArtFID (abbreviated as Score and FID respectively), and stylization time comparisons with recent 3D stylization models. Stylos achieves consistently favorable metric scores across the four scenes. Additionally, we follow StyleID ( Chung et al., 2024 ) and calculate additional metrics as reported in A.4 Table 5 and Table 6 .
Figure 5: Visual comparisons between Stylos and recent 3D stylization approaches. Stylos successfully transfers diverse artistic styles to the scenes while preserving fine structural details.
Figure 6: Evaluation of multi-style blending and controllable stylization. Top: interpolation between two style embeddings. Bottom: interpolation between the content embedding and a style embedding.
Appendix figures & tables21 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Scaling of inference time and GPU memory with the number of views per batch. Time is averaged over 100 iterations, and memory is reported as the peak allocation, without rendering cost.
Figure 8: Visual comparisons of different style losses on the CO3D skateboard scene. It is clear that the 3D style loss provides a stronger sense of 3D geometry and cleaner wooden floor textures evaluated on varying style images.
Figure 9: Visual comparisons of different style losses on the CO3D pizza scene. It is clear that the 3D style loss provides stronger artistic stylization to the table surface, while the image style loss often fails to impose the expected styles to cover the whole table surface.
Figure 10: Visual comparisons of different style losses on the CO3D donut scene. It is clear that the 3D style loss provides faithful stylization to the table cover while preserving the seams on the tablecloth compared to both the image style loss and the scene style loss.
Method
Train
Truck
FID ↓
LPIPS ↓
Gray ↓
CFSD ↓
CM Loss ↓
FID ↓
LPIPS ↓
Gray ↓
CFSD ↓
CM Loss ↓
StyleGaussian
34.59
0.483
0.377
0.190
0.418
28.53
0.522
0.367
0.131
0.382
G-Style
14.80
0.471
0.364
0.165
0.185
13.65
0.512
0.353
0.087
0.176
StylizedGS
22.90
0.707
0.648
0.216
0.396
22.65
0.779
0.713
0.084
0.417
SGSST
24.15
0.520
0.409
0.220
0.257
19.50
0.577
0.445
0.177
0.196
Styl3R
19.77
0.670
0.598
0.244
0.364
19.28
0.682
0.575
0.102
0.350
Appendix
Table 5: Following StyleID ( Chung et al., 2024 ) , we additionally report sylization quality metrics, FID, LPIPS, LPIPS-gray, CFSD, color matching loss (HistoGAN loos), as supplementary to Table 4 .
Method
M60
Garden
FID ↓
LPIPS ↓
Gray ↓
CFSD ↓
CM Loss ↓
FID ↓
LPIPS ↓
Gray ↓
CFSD ↓
CM Loss ↓
StyleGaussian
30.54
0.506
0.413
0.138
0.467
25.23
0.569
0.454
0.189
0.480
G-Style
13.93
0.498
0.395
0.092
0.208
16.17
0.500
0.364
0.103
0.179
StylizedGS
28.33
0.815
0.728
0.102
0.443
33.73
0.876
0.827
0.074
0.570
SGSST
23.95
0.552
0.458
0.204
0.264
20.93
0.529
0.415
0.233
0.228
Styl3R
17.14
0.646
0.573
0.124
0.314
22.38
0.637
0.556
0.097
0.335
Appendix
Table 6: Following StyleID ( Chung et al., 2024 ) , we additionally report sylization quality metrics, FID, LPIPS, LPIPS-gray, CFSD, color matching loss (HistoGAN loos), as supplementary to Table 4 .
Figure 11: Failures caused by inaccurate scene reconstruction, such as missing geometric details or incorrect region colors. Top row: The input style image and the stylized multi-view results; mid row: a reference frame from the input contents and the reconstructed multi-view images; bottom row: the input contents.
Figure 12: Failures caused by generating colors that are absent from the original style images. Top row: The input style image and the stylized multi-view results; mid row: a reference frame from the input contents and the reconstructed multi-view images; bottom row: the input contents.
Figure 13: Failures caused by introducing blurry or over-smoothed style colors. Top row: The input style image and the stylized multi-view results; mid row: a reference frame from the input contents and the reconstructed multi-view images; bottom row: the input contents.
Figure 14: Style images from WikiArt (2025) used for model evaluation in all quantitative comparisons, while not used for training Stylos.
Figure 15: Style images from WikiArt (2025) used for model evaluation in all quantitative comparisons, while not used for training Stylos.
Figure 16: Additional visual results of Stylos on diverse style–content pairs on Tanks and Temples Dataset ( Knapitsch et al., 2017 ) . Stylos consistently produces visually pleasing stylizations that preserve content geometry while transferring the target style.
Figure 17: Additional visual results of Stylos on the CO3D dataset. The top left is the style image while the rest of the top row are the rendered images. The mid row are the rendered images conditioned on the mid left imamge. The bottom row is the content inputs.
Figure 18: Additional visual results of Stylos on diverse style–content pairs.
Figure 19: Additional visual results of Stylos on diverse style–content pairs.
Figure 20: Additional visual results of Stylos on diverse style–content pairs.
Figure 21: Additional visual results of Stylos on diverse style–content pairs.
Figure 22: Additional visual results of Stylos on diverse style–content pairs.
Figure 23: Additional visual results of Stylos on the DL3DV-10K dataset. The top left is the style image while the rest of the top row are the rendered images. The mid row are the rendered images conditioned on the mid left imamge. The bottom row is the content inputs.
Figure 24: Additional visual results of Stylos on diverse style–content pairs.
Figure 25: Additional visual results of Stylos on diverse style–content pairs.
Artistic style transfer is well studied for images and videos, but extending it to multi-view 3D scenes remains difficult because stylization can disrupt correspondences needed by geometry-aware pipelines. Independent per-view stylization often causes texture drift, warped edges, and inconsistent shading, degrading SLAM, depth prediction, and multi-view reconstruction. This thesis addresses multi-view stylization that remains usable for downstream 3D tasks without assuming camera poses or an explicit 3D representation during training. We introduce a feed-forward stylization network trained with per-scene test-time optimization under a composite objective coupling appearance transfer with geometry preservation. Stylization is driven by an AdaIN-inspired loss from a frozen VGG-19 encoder, matching channel-wise moments to a style image. To stabilize structure across viewpoints, we propose a correspondence-based consistency loss using SuperPoint and SuperGlue, constraining descriptors from a stylized anchor view to remain consistent with matched descriptors from the original multi-view set. We also impose a depth-preservation loss using MiDaS/DPT and use global color alignment to reduce depth-model domain shift. A staged weight schedule introduces geometry and depth constraints. We evaluate on Tanks and Temples and Mip-NeRF 360 using image and reconstruction metrics. Style adherence and structure retention are measured by Color Histogram Distance (CHD) and Structure Distance (DSD). For 3D consistency, we use monocular DROID-SLAM trajectories and symmetric Chamfer distance on back-projected point clouds. Across ablations, correspondence and depth regularization reduce structural distortion and improve SLAM stability and reconstructed geometry; on scenes with MuVieCAST baselines, our method yields stronger trajectory and point-cloud consistency while maintaining competitive stylization.
Shirsha Bose
SCHOOL OF COMPUTATION, INFORMATION AND TECHNOLOGY - INFORMATICS TECHNISCHE UNIVERSITÄT MÜNCHEN
While 3D Gaussian Splatting (3DGS) provides an efficient and explicit representation for novel view synthesis, enforcing stylistic coherence across viewpoints remains challenging. Existing 3D stylization methods typically apply 2D feature-matching losses independently per rendered view, which leads to unstable style allocation, many-to-one feature reuse, and limited cross-view consistency. We propose a capacity-controlled framework for multi-view stylization of 3DGS, grounded in optimal transport. Specifically, we reformulate local style matching as a semi-balanced optimal transport problem. By introducing explicit column-capacity constraints with tunable strength, our formulation mitigates many-to-one matching and enables controllable allocation of style features. This transport-based objective provides a principled mechanism for balancing feature coverage and stylistic diversity while maintaining stable correspondences across viewpoints. To further enhance cross-view coherence, we incorporate a novel cross-view matching guidance to constrain correspondences between scene content and style patterns. In addition, we introduce several geometric regularizations to enhance the vanilla 3DGS, thereby enabling optimized Gaussian primitives to represent finer-grained textures during stylization. Extensive experiments demonstrate that our approach significantly improves multi-view stylistic consistency and produces stable, expressive 3D stylizations while preserving the core semantic structure of the scene.
Zhihao Wen, Yixin Yang, Bojian Wu +4
Guangdong Provincial Key Laboratory of Visual Media and Multidimensional Intelligence, CSSE, Shenzhen University, China · Tencent Games, China · The Hebrew University of Jerusalem, Israel +1
In this work, we focus on zero-shot 3D style transfer that can generate multi-view consistent stylized views of the 3D scene given an arbitrary style image. We primarily tackle the issue of data scarcity in 3D style transfer, which arises when each model is trained on only a single scene, thereby limiting the number of available content images. This scarcity significantly hampers stylization performance, as model optimization relies on a sufficient number of content-style image pairs to provide supervisory signals. Our core idea is to integrate a decoder pre-trained on large-scale 2D image datasets into the 3D style transfer pipeline, thereby leveraging the prior knowledge encoded in the decoder from learning over numerous content-style image pairs. Our method combines feature Gaussian splatting and deferred stylization, enabling high-quality stylization with the data-sufficient decoder network while ensuring view consistency by unifying view-dependent operations into a view-invariant process. Experiments demonstrate that our Data-Sufficient StyleGaussian (DS-StyleGaussian) model outperforms existing zero-shot 3D style transfer methods in terms of visual quality across various datasets. This work also suggests that 2D pre-training can serve as a strong enhancement for 3D tasks, bridging the data gap between 2D and 3D.
Xin Dong, Yunzhi Teng, Wenfeng Deng +1
Shenzhen International Graduate School, Tsinghua University · Pengcheng Laboratory