cs.CVApr 9, 2026

Image-Guided Geometric Stylization of 3D Meshes

Authors: Changwoon Choi, Hyunsoo Lee, Clément Jambon, Yael Vinker, Young Min Kim

Organizations: Seoul National University · Massachusetts Institute of Technology (MIT)

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

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. Project page: https://changwoonchoi.github.io/GeoStyle

Figures & tables

Figure 2: Text-based deformation often struggles to capture and transfer geometric style, whereas our image-guided framework can transfer intended aesthetic from rich and specific visual cues.
Figure 3: Method overview. We first extract geometric style from reference images by training LoRA using a DreamBooth-style objective ( Sec. 3.1.1 ). We optimize per-face Jacobians to transfer geometric style to source mesh. The optimization is guided by SDS loss ( Sec. 3.1.2 ). To handle both large structural deformation and fine-grained details, we adapt a coarse-to-fine deformation ( Sec. 3.3 ), along with the cage-based regularization ( Sec. 3.2.1 ) and optional symmetry regularization ( Sec. 3.2.2 ).
Figure 4: Effect of model choices on text-guided mesh deformation. For each source mesh, we apply SDS [ 46 ] guidance using (1) DeepFloyd-IF [ 3 ] , (2) SDXL [ 45 ] with the original VAE [ 22 ] , and (3) SDXL with our approximated encoder. Our setup successfully transfers the intended semantic shape, while the alternatives either distort geometry or fail to deform the global structure.
Figure 5: Qualitative comparison of the proposed method against baselines [ 35 , 20 , 21 , 41 , 8 ] . 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, whereas baselines struggle to capture the intended geometry.
Figure 6: User study results. We examine each method using three criteria, including the measurement of geometry alignment, content preservation, and aesthetic style transfer.
Figure 7: Qualitative geometric stylization results using diverse source mesh and style references. Our method effectively transfers the geometric style from the reference image while preserving the overall semantics of the source mesh.
Figure 8: Ablation study results. 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.
Figure 9: Geometric stylization results with text conditioning. Our method effectively incorporates textual prompts alongside style reference images. The resulting meshes align with the text-described contents while consistently maintaining the geometric style encoded in the style reference.
Figure 10: Our approach can transfer geometric style only to user-selected regions while preserving the original shape elsewhere.
Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

MethodFID ( ↓\downarrow )
Paparazzi [ 35 ]419.89
Neural 3D Mesh Renderer [ 20 ]417.08
MeshUp [ 21 ]396.23
Text2Mesh [ 41 ]409.17
TextDeformer [ 8 ]400.58
Ours376.57
Appendix
Table 1: Quantitative comparison measured by FID [ 13 ] .
Figure 11: Additional qualitative comparison of our method with the baselines [ 35 , 20 , 21 , 41 , 8 ] on diverse geometric stylization scenarios.
Figure 12: Additional qualitative results of the proposed method.
More in this topic

3D Mesh Deformation

See all 16 papers

Explore similar work

CardsList

  1. DreamStyle3D: Efficient 3D Stylized Asset Generation via Dual-Attention Disentanglement

    Jul 27, 2026Kai Wang, Ziheng Ouyang, Xuying Zhang +2Disentangled Representation Learning3D Asset Generation

    VCIP, Nankai University Tianjin, China · Zhongguancun Academy Beijing, China · Joy Future Academy, JD Beijing, China +1

  2. InStyle: Instant Appearance Stylization of 3D Shapes

    Dec 3, 2025Melis Ocal, Xiaoyan Xing, Yue Li +33D Shape Editing3D Representation Learning

    University of Amsterdam · Bosch Center for AI

  3. AnyStyle: Single-Pass Multimodal Stylization for 3D Gaussian Splatting

    Feb 3, 2026Joanna Kaleta, Bartosz Świrta, Kacper Kania +33D Gaussian SplattingFeed-Forward 3D Reconstruction

    Warsaw University of Technology · Sano Centre for Computational Medicine · IDEAS NCBR +2