← ArchivePaper2018
3D Hair Synthesis Using Volumetric Variational Autoencoders
Represents the 3D hairstyle manifold via a volumetric VAE trained on orientation fields; synthesizes new hairstyles from a single image in one second.
Abstract
Recent advances in single-view 3D hair digitization have made the creation of high-quality CG characters scalable and accessible to end-users, enabling new forms of personalized VR and gaming experiences. To handle the complexity and variety of hair structures, most cutting-edge techniques rely on the successful retrieval of a particular hair model from a comprehensive hair database. Not only are the aforementioned data-driven methods storage intensive, but they are also prone to failure for highly unconstrained input images, complicated hairstyles, and failed face detection. Instead of using a large collection of 3D hair models directly, we propose to represent the manifold of 3D hairstyles implicitly through a compact latent space of a volumetric variational autoencoder (VAE). This deep neural network is trained with volumetric orientation field representations of 3D hair models and can synthesize new hairstyles from a compressed code. To enable end-to-end 3D hair inference, we train an additional embedding network to predict the code in the VAE latent space from any input image. Strand-level hairstyles can then be generated from the predicted volumetric representation. Our fully automatic framework does not require any ad-hoc face fitting, intermediate classification and segmentation, or hairstyle database retrieval.
How to read this
- Category
- Method: deep generative model for single-image 3D hair synthesis
- Contributions
- Represents the manifold of 3D hairstyles implicitly through the compact latent space of a volumetric VAE trained on orientation-field representations
- Adds an embedding network that predicts the latent code from any input image for end-to-end 3D hair inference
- Generates strand-level hairstyles from the predicted volumetric representation, reportedly from a single image in about one second
- Context
- Departs from retrieval-based single-view hair digitization (e.g. AutoHair: Fully Automatic Hair Modeling from a Single Image) by replacing a large 3D hair database with a learned, compact latent space.Builds on: AutoHair: Fully Automatic Hair Modeling from a Single Image
- Correctness
- Assumes the volumetric orientation-field VAE captures enough of the hairstyle manifold to generalize beyond retrieval; this should reduce storage and the failure modes of database lookup (unconstrained images, failed face detection), but quality is bounded by the training data and the volumetric-to-strand conversion, so very complex or out-of-distribution styles remain a question.
- Clarity
- Requires familiarity with VAEs and volumetric fields; a first pass conveys the pipeline, a second pass is needed for the representation and training.
- How to read it
- First pass for the retrieval-versus-latent-space framing and the orientation-field representation; second pass on the VAE and embedding networks if building generative hair or single-image reconstruction.
Related work
Keywords
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 →