Geometric Stylization of 3D Representations

42 concepts34 papers

Early style transfer algorithms treated artistic style as an additive texture—a coat of paint applied over a fixed geometric structure. However, as noted by Zheng Xu et al. in Beyond Textures, this 2D texture-matching paradigm fails to capture artistic styles defined by volumetric properties, lighting, and spatial depth. True artistic stylization requires altering the geometry itself: a cubist sculpture doesn't just have a cubist texture, its physical form is broken into planes; a cartoon character doesn't just have flat shading, its proportions are structurally deformed.

The field of geometric stylization bridges this gap by treating style as a joint transformation of appearance and physical structure. By optimizing 3D representations—such as meshes, neural radiance fields, or Gaussian splats—to conform to geometric style priors, researchers can fundamentally alter an object's shape while preserving its underlying semantic content.

Sparsity and Optimization on Explicit Meshes

How do you mathematically force a smooth 3D mesh to adopt a specific geometric style, like cubism, without destroying its recognizable features? Hsueh-Ti Derek Liu et al. provide an elegant optimization-based answer in Cubic Stylization.

The intuition relies on surface normals. Smooth objects have surface normals that vary continuously across space. Sharp, polyhedral objects (like cubes) have normals that are highly clustered—they point only in a few discrete directions.

Smooth Surface:   Normals vary continuously ( ↗ ↑ ↖ → )
Cubic Surface:    Normals are sparse/discrete ( ↑ ↑ → → )

By applying an $\ell_1$-regularization to rotated surface normals, the optimization penalizes variation and forces the normals to become sparse, effectively flattening the mesh faces into sharp planes. To prevent the mesh from collapsing into a single point or tearing apart, the authors couple this sparsity loss with an "as-rigid-as-possible" (ARAP) energy term. This acts as a geometric anchor, ensuring that local patches deform but retain their original structural relationships.

Rearranging Primitives in Volumetric and Radiance Fields

When moving from explicit meshes to implicit or point-based representations, the optimization variables change from vertices to densities, colors, or spatial primitives. The authors of SNeRF demonstrated that authentic 3D stylization requires modifying both volumetric density and color simultaneously, preventing the scene from looking like a flat texture projected onto a rigid model.

With the advent of 3D Gaussian Splatting (3DGS), researchers gained a highly flexible representation where geometry is explicitly defined by millions of physical primitives (Gaussians). Dmytro Kotovenko et al. leverage this in WaSt-3D by framing stylization as an optimal transport problem.

Instead of merely changing the color of existing Gaussians, they use the Sinkhorn divergence (a debiased Wasserstein-2 distance) to match the spatial distribution of a content scene with the geometric primitives of a style scene. The algorithm physically rearranges style-specific building blocks (e.g., the individual shapes of pebbles, brushstrokes, or grass blades) to assemble the macro-geometry of the target content. The style is not painted onto the shape; the shape is built out of the style.

Guiding Structural Deformation with Generative Priors

To achieve more complex, semantic deformations—such as turning an organic shape into a mechanical one—optimization metrics like surface normal sparsity or primitive distribution matching are insufficient. The system needs a semantic understanding of the target style's geometry.

Recent frameworks integrate large generative models to provide this structural guidance:

  • Text-Driven Deformations: Jamie Wynn et al. introduce Morpheus, which uses text-driven RGBD diffusion models to guide multi-view consistent geometric deformations in 3D Gaussian Splats, allowing abstract or structural transformations based purely on textual descriptions.
  • Attention-Level Guidance: Yitong Yang et al. propose DiffStyle3D, which injects explicit geometric guidance directly into the self-attention layers of a diffusion model. This eliminates multi-view flickering during structural deformation by ensuring the model's attention maps are anchored to the underlying 3D geometry.
  • Disentangling Texture and Geometry: For zero-shot asset generation, Zefan Qu et al. develop StyleSculptor. They utilize a dual-guidance mechanism that disentangles texture from geometry, allowing intricate geometric motifs (like relief patterns or sharp architectural edges) to be transferred without leaking the source content's identity.
  • Direct Primitive Optimization: On a universal scale, Kornel Howil demonstrates in CLIPGaussian that directly optimizing the physical shapes and positions of Gaussian primitives via multimodal CLIP guidance can capture complex structural textures like thick paint impastos or relief-like patterns without requiring massive generative networks.

Vectorized Primitives and Contours

Another way to break free from rigid geometry is to change the underlying representation entirely during rendering. In image-space rendering, line drawings often suffer from broken outlines because the model lacks 3D awareness, a limitation discussed by Pierre Bénard et al. in Line Drawings from 3D Models.

To solve this, Difan Liu et al. introduce Neural Strokes, a differentiable rendering framework that explicitly models stroke thickness and geometric deformation based on the underlying 3D shape. By operating on vector primitives rather than pixels, it preserves fine geometric details while allowing expressive, non-photorealistic line deformations.

Taking this a step further, Hao-Bin Duan et al. propose Neural 3D Strokes, which abandons traditional dense volumes or meshes entirely. Instead, they represent the 3D scene using vectorized 3D splines. Because the primitives themselves are strokes with physical width and curvature, the scene inherently possesses structural artistic properties. This allows for perfect view-consistency and radical geometric stylization that surface-level color adjustments could never achieve.

Key Foundational Work

Beyond Textures: Learning from Multi-domain Artistic Images for Arbitrary Style Transfer (2018) — contributed 1 concept elements to this direction.
Semi-Supervised Co-Analysis of 3D Shape Styles from Projected Lines (2018) — contributed 1 concept elements to this direction.
Cubic Stylization (2019) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_geometric_stylization_9bd3c888,
  title = {geometric stylization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/geometric-stylization/rdr_9bd3c88800e64c99a5348752ad312d92},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_geometric_stylization_9bd3c888_v1,
  title = {geometric stylization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/geometric-stylization/rdr_9bd3c88800e64c99a5348752ad312d92/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}

See something wrong? Report a mistake