What is TANGLED
TANGLED is a 3D hairstyle generation method jointly launched by ShanghaiTech University, Deemos Technology, and Huazhong University of Science and Technology. It supports generating high-quality 3D hair strands from images of any style and viewpoint. TANGLED is implemented based on three core steps: using the diverse MultiHair dataset to provide rich hairstyle samples; a diffusion framework based on multi-view line drawings, using cross-attention of line drawing features to capture the topological structure of hairstyles; and a parametric post-processing module to repair the details of complex hairstyles (such as braids). TANGLED improves the realism and diversity of hairstyles, supports the creation of culturally inclusive digital characters, and provides new application possibilities for fields such as animation and augmented reality.
Main Features of TANGLED
- Diverse Hairstyle Generation: Supports handling various complex hairstyles, such as braids, curly hair, traditional hairstyles, etc.
- Multi-view Input Support: Accepts single-view or multi-view images as input.
- Flexible Input Styles: Supports multiple input styles, including photos, hand-drawn sketches, cartoons, oil paintings, etc., meeting the needs of different application scenarios.
- Cultural Inclusivity: Particularly focuses on underrepresented hairstyle textures (such as curls and braids) and complex geometric structures, supporting the generation of culturally significant hairstyles.
- Efficient Integration: The generated 3D hairstyles can be directly integrated into existing CG pipelines, such as Blender and Unreal Engine.
Technical Principles of TANGLED
- NeuraPressMultiHair Dataset: Provides 457 diverse hairstyles, annotated with 74 attributes, with a focus on complex and culturally significant hairstyles. The dataset enhances diversity based on multi-view rendering and line drawing extraction, and uses GPT-4 to generate detailed text annotations.
- Diffusion Framework: A diffusion model based on multi-view line drawings. Line drawing features capture the topological structure of hairstyles (such as hair strand density and parting lines). DINOv2 is used to extract line drawing features, which are integrated into the diffusion model based on a cross-attention mechanism. Line drawing features from different viewpoints are randomly mixed to enhance the model's adaptability to different input styles and viewpoints.
- Parametric Post-processing Module: Parametric modeling and repair techniques are designed for complex hairstyles (such as braids). The Frenet-Serret framework is used to generate the geometric structure of braids, which are naturally integrated into the hairstyle based on an attention mechanism. Laplacian smoothing technology is used to reduce high-frequency noise, ensuring the geometric coherence and visual effect of the hairstyle.
Project Address of TANGLED
- Project Official Website：https://sites.google.com/view/tangled
- arXiv Technical Paper：https://arxiv.org/pdf/2502.06392v1
Application Scenarios of TANGLED
- Culturally Inclusive Virtual Character Creation: Generates diverse hairstyles, supporting virtual character design for different cultural backgrounds.
- Animated Hairstyle Design: Generates 3D hairstyles from sketches, improving animation production efficiency.
- Augmented Reality (AR) Hairstyle Preview: Combines AR technology to allow users to virtually try on different hairstyles.
- Virtual Makeup Try-on Applications: Helps users preview hairstyle effects before purchasing.
- Sketch-based 3D Hair Strand Editing: Quickly adjusts hairstyles based on modified sketches, suitable for the creative design field.