Research map / 3D vision / 3D generation
Text-Conditioned Generative Model of 3D Strand-based Human Hairstyles
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.
Proposes HAAR, a latent diffusion model that generates strand-based 3D hairstyles from text, trained on VQA-annotated synthetic hairstyles in a UV hair-map latent space.
Problem
Text-to-3D hair generation methods based on SDS only produce outer hair surfaces without internal strand structure, so outputs are not usable for graphics simulation or physics-based rendering.
Approach
Encode strand-based hairstyles as latent hair maps on the scalp via a VAE, train a text-conditioned EDM latent diffusion model on VQA-generated captions of synthetic hairstyles, and upsample guiding strands with a blend of nearest-neighbor and bilinear interpolation in latent space.
Key results
- HAAR achieves FID 36.9 vs 75.5 for Neural Haircut in unconditional comparison (Table 1).
- In a user study with over 1,200 responses, HAAR was preferred over Neural Haircut in 87.5% of comparisons.
- HAAR generates a hairstyle in about 4.3 seconds, versus hours for SDS-based TECA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add cosine noise scheduling (ablation) | HAAR · Unconditional hairstyle generation | FID: 62.1 -> 36.9 | Helped |
| Add SNR weighting (ablation) | HAAR · Unconditional hairstyle generation | FID: 44.7 -> 36.9 | Helped |
| Use frontal and back views rather than frontal only (ablation) | HAAR conditioning annotations · Text-to-hairstyle semantic matching | CSIM: 0.186 -> 0.189 | Helped |
| Aggregate answers from multiple questions rather than one question (ablation) | HAAR conditioning annotations · Text-to-hairstyle semantic matching | CSIM: 0.18 -> 0.189 | Helped |
| Train on 10% rather than the full training samples (ablation) | HAAR · Text-to-hairstyle semantic matching | CSIM: 0.189 -> 0.187 | Hurt |
| Train on 200% rather than the full training samples (ablation) | HAAR · Text-to-hairstyle semantic matching | CSIM: 0.189 -> 0.191 | Helped |
| Blend nearest-neighbor and bilinear interpolation in latent space instead of using either alone (ablation) | HAAR upsampling · Hairstyle upsampling | visual realism | Helped |
| Add noise to interpolated latents (ablation) | HAAR upsampling · Hairstyle upsampling | visual realism | Helped |
| Use HAAR instead of Neural Haircut | Unconditional hairstyle generation · Unconditional hairstyle generation | FID: 75.5 -> 36.9 | Helped |
| Use HAAR instead of Neural Haircut | Unconditional hairstyle generation · Unconditional hairstyle generation | MMD: 31507.7 -> 21104.9 | Helped |
Surprising findings
- Off-the-shelf VQA systems such as LLaVA were inaccurate on hair descriptions, so a custom multi-question annotation pipeline was needed.
Limitations
- Quality limited by dataset variety and accuracy of VQA-generated annotations.
- Scalp interpenetration artifacts and insufficient curliness for extreme styles such as afro.
- Only geometry is generated; hair color and texture are not modeled.
Experimental setup
Datasets: artist-created hairstyle dataset (40 hairstyles) CT2Hair USC-HairSalon
Benchmarks: Neural Haircut training distribution comparison text-hairstyle CLIP similarity
Metrics: MMD COV 1-NNA FID CSIM user preference
Baselines: TECA Neural Haircut NH++
Builds on: Neural Haircut EDM Min-SNR weighting Imagic
Closest papers by meaning
- Doubly Hierarchical Geometric Representations for Strand-based Human Hairstyle Generation (NeurIPS 2024 · similarity 0.91)
- Perm: A Parametric Representation for Multi-Style 3D Hair Modeling (ICLR 2025 · similarity 0.90)
- DiffLocks: Generating 3D Hair from a Single Image using Diffusion Models (CVPR 2025 · similarity 0.89)
- MonoHair: High-Fidelity Hair Modeling from a Monocular Video (CVPR 2024 · similarity 0.88)
- SimAvatar: Simulation-Ready Avatars with Layered Hair and Clothing (CVPR 2025 · similarity 0.88)
- DGH: Dynamic Gaussian Hair (NeurIPS 2025 · similarity 0.87)
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.