Image-Guided Geometric Stylization of 3D Meshes

CVPR 2026
1 Seoul National University 2 MIT
(* Denotes equal contribution)
1 2

Given a 3D mesh and style reference images, we transfer geometric style — poses, proportions, or distinctive geometric traits, while preserving the original mesh's topology and semantic identity.

Abstract

Recent generative models can create visually plausible 3D representations of objects. However, the generation process often allows for implicit control signals, such as contextual descriptions, and rarely supports bold geometric distortions beyond existing data distributions. We propose a geometric stylization framework that deforms a 3D mesh, allowing it to express the style of an image. While style is inherently ambiguous, we utilize pre-trained diffusion models to extract an abstract representation of the provided image. Our coarse-to-fine stylization pipeline can drastically deform the input 3D model to express a diverse range of geometric variations while retaining the valid topology of the original mesh and part-level semantics. We also propose an approximate VAE encoder that provides efficient and reliable gradients from mesh renderings. Extensive experiments demonstrate that our method can create stylized 3D meshes that reflect unique geometric features of the pictured assets, such as expressive poses and silhouettes, thereby supporting the creation of distinctive artistic 3D creations.

Style Transfer in 3D

Following the success of 2D image style transfer, style transfer has been extended to 3D across various representations such as point clouds, meshes, NeRFs, and 3D Gaussian Splatting. Most of these works, however, are focusing on transferring high-frequency patch statistics of colors and textures. A parallel line of mesh-deformation methods does modify shape, but each minimizes a hand-crafted energy function tailored to a specific aesthetic (e.g., cubic stylization), and thus supports only a narrow class of styles.

Majority of style transfer methods in 3D focus on transferring high-frequency patch statistics of colors and textures.

Style is not only about color or texture. What truly distinguishes an artistic 3D creation often lies in its geometry: the blocky facets of a cubic sculpture, the balloon-like proportions of an inflated toy, the elongated limbs of Bourgeois's spider, or the expressive pose of a figure. Such traits are beyond the reach of patch statistics, and also resist unambiguous description through text.

Geometric style goes beyond surface texture — cubic sculptures, balloon-like proportions, elongated spider limbs, and expressive poses all define style through geometry.

Our framework takes a source 3D mesh and style reference images as input, and deforms the mesh to reflect their geometric style. The key idea is to use a personalized diffusion model, fine-tuned on the style reference images, as a style-specific optimizer that guides mesh deformation — covering a far broader range of styles than any hand-crafted energy can express.

Method

Our geometric stylization framework takes a source mesh and style reference images as input, and deforms the mesh to exhibit the geometric style depicted in the images through a two-stage approach: (1) style extraction and (2) mesh deformation.

  • Stage 1: Geometric style extraction We extract geometric style from style references by training LoRA weight using DreamBooth, to capture both global shape and structural characteristics.
  • Stage 2: Mesh deformation
    • Coarse deformation. Using the pre-trained SDXL with an approximated VAE encoder, we compute SDS loss to induce large but semantically consistent deformations. A cage-based deformation technique driven by an auxiliary mesh enables coarse structure changes. Deformation is optionally enhanced with symmetry regularization.
    • Fine deformation. To refine fine-grained geometric details while preserving overall structure, we use additional stage without the cage-based deformation, also guided by SDS loss.

Interactive Results

Source Mesh

Style Image

Style reference

Deformed Mesh

Drag to rotate · Scroll to zoom · Meshes may take a few seconds to load

Comparison with Other Methods

Our method achieves expressive geometric deformation, accurately reflecting both coarse structure and fine detail from the style reference while preserving the identity of source mesh, baselines struggle to capture the intended geometry.

Additional Results

Geometric Stylization with Text Condition

Our method can incorporate textual prompts alongside style reference images. The resulting meshes align with the text-described contents while maintaining the geometric style encoded in the style reference.

Localized Deformations

Our approach can transfer geometric style only to user-selected regions while preserving the original shape elsewhere.

Ablation Study

We analyze the effects of the approximated VAE encoder (1st row), cage-coefficient regularization (2nd) and symmetry loss (3rd). The qualitative results show that each component is essential for producing high-fidelity stylization result.

References

[1] Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. Image Style Transfer Using Convolutional Neural Networks. CVPR 2016.
[2] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3D Mesh Renderer. CVPR 2018.
[3] Hsueh-Ti Derek Liu, Michael Tao, and Alec Jacobson. Paparazzi: Surface Editing by Way of Multi-View Image Processing. ACM TOG 2018.
[4] Hsueh-Ti Derek Liu and Alec Jacobson. Cubic Stylization. ACM TOG 2019.
[5] Lukas Höllein, Justin Johnson, and Matthias Nießner. StyleMesh: Style Transfer for Indoor 3D Scene Reconstructions. CVPR 2022.
[6] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2Mesh: Text-Driven Neural Stylization for Meshes. CVPR 2022.
[7] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. ARF: Artistic Radiance Fields. ECCV 2022.
[8] William Gao, Noam Aigerman, Thibault Groueix, Vladimir G. Kim, and Rana Hanocka. TextDeformer: Geometry Manipulation Using Text Guidance. SIGGRAPH 2023.
[9] Hyunwoo Kim, Itai Lang, Noam Aigerman, Thibault Groueix, Vladimir G. Kim, and Rana Hanocka. MeshUp: Multi-Target Mesh Deformation via Blended Score Distillation. 3DV 2025.

BibTeX

Click to copy
@inproceedings{choi2026geostyle,
  title={Image-Guided Geometric Stylization of 3D Meshes},
  author={Choi, Changwoon and Lee, Hyunsoo and Jambon, Clément and Vinker, Yael and Kim, Young Min},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}