# Text2Mesh: Text-Driven Neural Stylization for Meshes

**Authors:** Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, Rana Hanocka  
**arXiv:** [2112.03221](https://arxiv.org/abs/2112.03221)  
**Date:** 2021-12-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2112.03221v1/

## Key points

- Text2Mesh is a framework for controlling and editing the style of 3D objects
- It uses a disentangled representation of a 3D object with a fixed mesh input and a neural network called the neural style field network
- The framework can handle low-quality meshes without UV parameterization or specialized datasets
- It can generate stylizations over different types of 3D meshes, including outerwear variations, muscle definition, and hair details
- Text prompts are used as an easily modifiable and expressive means to control style
- The technique extends beyond text-driven stylization to other modalities like images and 3D meshes using CLIP (Contrastive Language-Image Pretraining)
- Text2Mesh enables semantic manipulation of style for 3D shapes while preserving global semantics and underlying content
- Experiments were conducted to evaluate the method's performance, comparing it to baseline methods through user studies
- The paper discusses limitations in their approach

## Summary

In this work, the authors present Text2Mesh, a framework for intuitive control and editing of the style of 3D objects. The framework utilizes a disentangled representation of a 3D object using a fixed mesh input coupled with a neural network called the neural style field network. This network predicts color and local geometric details that conform to a target text prompt, allowing for the stylization of 3D meshes. One notable feature of Text2Mesh is its ability to handle low-quality meshes with arbitrary genus without requiring UV parameterization or specialized 3D mesh datasets. The framework can generate stylizations over a wide variety of 3D meshes, synthesizing different styles such as outerwear variations, muscle definition and hair details. The authors emphasize the use of text prompts as an easily modifiable and expressive means to control style. They demonstrate that their technique extends beyond text-driven stylization to other target modalities like images and 3D meshes by harnessing the representational power of CLIP (Contrastive Language-Image Pretraining). Text2Mesh enables semantic manipulation of style for 3D shapes. The paper also discusses related work in text-driven manipulation and highlights how their method differs from existing techniques. They provide examples showcasing the fine-grained controls offered by their approach, generating high-fidelity details while preserving global semantics and underlying content. The authors conduct experiments to evaluate their method's performance and explore different aspects such as learning color and geometry together, comparing it to baseline methods through user studies. They also discuss limitations in their approach. Overall, Text2Mesh presents an innovative technique for semantic manipulation of style in 3D meshes using text prompts and CLIP-guided neural networks. It offers intuitive controls for generating diverse stylizations across various types of 3D objects while maintaining high fidelity and handling low-quality meshes.