Research map / 3D vision / Face and shape
Geometry in Style: 3D Stylization via Surface Normal Deformation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Geometry in Style stylizes triangle meshes from text prompts by optimizing target vertex normals, mapped to deformations by a single-step differentiable ARAP layer (dARAP) and guided by Cascaded Score Distillation.
Problem
Text-guided 3D mesh stylization that makes expressive, detailed deformations while preserving the identity of the source shape, which existing bump-map or Jacobian-based methods fail to balance.
Approach
Represent a mesh deformation as per-vertex target normals; a differentiable ARAP layer (one local Procrustes rotation step plus one global Poisson solve) converts normals into deformed vertices, and a text-conditioned diffusion loss (CSD with DeepFloyd IF) optimizes the normals.
Key results
- Triangle area ratio of 1.080 (std 0.233) for Geometry in Style vs 1.288 (std 0.363) for MeshUp and 0.827 (std 0.360) for TextDeformer, indicating better identity preservation.
- Optimization runs 2500 epochs taking about 2 hours 15 minutes on a single A40 GPU.
- Single dARAP local-global step suffices within the optimization pipeline, avoiding iterative ARAP to convergence.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use target-normal deformation with dARAP instead of TextDeformer | Geometry in Style · 20 mesh-prompt pairs | triangle area ratio mean closeness to 1: 0.827 -> 1.080 | Helped |
| Use target-normal deformation with dARAP instead of MeshUp | Geometry in Style · 20 mesh-prompt pairs | triangle area ratio mean closeness to 1: 1.288 -> 1.080 | Helped |
| Use target-normal deformation with dARAP instead of TextDeformer | Geometry in Style · 20 mesh-prompt pairs | triangle area ratio std. dev.: 0.360 -> 0.233 | Helped |
| Use target-normal deformation with dARAP instead of MeshUp | Geometry in Style · 20 mesh-prompt pairs | triangle area ratio std. dev.: 0.363 -> 0.233 | Helped |
| Increase inference λ from 5 to 8 to 11 (ablation) | Geometry in Style · Robot lion stylization | stylization strength | Helped |
| Decrease inference λ below the optimization value of 8 (ablation) | Geometry in Style · Robot lion stylization | stylization strength | Hurt |
Surprising findings
- A single ARAP local-global step is sufficient for high-quality deformations when used inside gradient descent, rather than iterating to convergence.
Limitations
- Relies on the cotangent Laplacian, which requires manifold meshes and is sensitive to triangle aspect ratios; inputs must be remeshed.
- Deformed meshes may self-intersect.
- Topology-preserving; cannot add parts or change topology.
Experimental setup
Metrics: triangle area ratio mean triangle area ratio standard deviation CLIP similarity
Baselines: TextDeformer MeshUp
Builds on: As-Rigid-As-Possible Neural Jacobian Fields Cascaded Score Distillation Liu and Jacobson normal-driven stylization
Closest papers by meaning
- LeGO: Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example (CVPR 2024 · similarity 0.89)
- Styl3R: Instant 3D Stylized Reconstruction for Arbitrary Scenes and Styles (NeurIPS 2025 · similarity 0.86)
- Implicit-ARAP: Efficient Handle-Guided Neural Field Deformation via Local Patch Meshing (NeurIPS 2025 · similarity 0.86)
- 3DToonify: Creating Your High-Fidelity 3D Stylized Avatar Easily from 2D Portrait Images (CVPR 2024 · similarity 0.85)
- Morpheus: Text-Driven 3D Gaussian Splat Shape and Color Stylization (CVPR 2025 · similarity 0.85)
- ARMesh: Autoregressive Mesh Generation via Next-Level-of-Detail Prediction (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.