← ArchivePaper2023

HAAR: Text-Conditioned Generative Model of 3D Strand-Based Human Hairstyles

Vanessa Sklyarova, Egor Zakharov, Otmar Hilliges, Michael J. Black, Justus Thies

Text-driven latent diffusion model generating dense strand-based hairstyles in UV space from VQA-annotated synthetic data, upsampling to 100K strands.

Open the paper ↗Trace in the lineage graph

Abstract

We present HAAR, the first text-conditioned generative model that produces classical 3D strand-based human hairstyles directly usable in computer graphics rendering and simulation pipelines. Unlike prior text-to-3D methods that rely on 2D priors and only recover the outer visible shell, HAAR represents a hairstyle as a UV texture map on the scalp whose pixels encode latent embeddings of individual strands learned by a VAE, and a latent diffusion model generates this map conditioned on a text embedding via cross-attention. To obtain training annotations, we render synthetic hairstyles and use an off-the-shelf visual question answering system together with a custom prompt pipeline to produce hairstyle descriptions. The generated guiding strands are upsampled in latent space by blending nearest neighbor and bilinear interpolation, yielding dense hairstyles in seconds rather than the hours required by SDS-based approaches, and the learned latent space also supports semantic editing of hairstyles from text.

How to read this

Category: Method: a text-conditioned generative model for strand-based hair
Contributions: A latent diffusion model that generates classical 3D strand-based hairstyles from text, directly usable in rendering and simulation pipelines
A UV-space scalp texture whose pixels encode per-strand latent embeddings learned by a VAE, conditioned on text via cross-attention
A VQA-plus-prompt pipeline to auto-annotate synthetic hairstyles, and a latent upsampling scheme producing 100K-strand grooms in seconds with text-driven editing
Context: Builds on prior-guided strand reconstruction (Neural Haircut) and hierarchical generative grooming (GroomGen), but replaces SDS-style 2D priors with a full 3D strand representation.Builds on: Neural Haircut: Prior-Guided Strand-Based Hair Reconstruction · GroomGen: A High-Quality Generative Hair Model Using Hierarchical Latent Representations
Correctness: Demonstrated on synthetic hairstyles annotated by an off-the-shelf VQA system, so output diversity and text fidelity are bounded by the synthetic training distribution and the quality of auto-generated captions; treat real-world coverage and caption accuracy as open.
Clarity: Accessible at a high level; a first pass conveys the UV-strand-plus-diffusion idea, a second pass is needed for the VAE encoding and latent upsampling math.
How to read it: First pass for the representation (UV map of strand latents) and the text-conditioning route; do a second pass on the VAE and upsampling if you care about how dense grooms stay coherent.

Builds on

Built upon by

Nothing yet.

Related work

Neural Haircut: Prior-Guided Strand-Based Hair Reconstruction 2023 / ICCV
Perm: A Parametric Representation for Multi-Style 3D Hair Modeling 2024 / ICLR
3D Hair Synthesis Using Volumetric Variational Autoencoders 2018 / SIGGRAPH Asia
Neural Strands: Learning Hair Geometry and Appearance from Multi-View Images 2022 / ECCV

Keywords

haar text-conditioned generation strand-based hairstyles latent diffusion model variational autoencoder hair texture map visual question answering vision-language model blip clip score distillation sampling cross-attention latent upsampling neural haircut teca hairstyle editing

This page summarises the entry and links to its original source. The archive never hosts or redistributes the publication itself.Show it in the full archive list →