Quaffure: Real-Time Quasi-Static Neural Hair Simulation

TL;DR

Quaffure tackles real-time, realistic hair drape across varied body poses and shapes by introducing a two-stage neural framework: a pose-based groom transformation and a learned groom deformation decoder, operating on a texture-based hair encoding. Training uses a physics-based self-supervision loss that combines gravity, body and self-collision, and a pose regularizer, with a two-term elastic potential built from a modified Cosserat energy to balance realism against training speed. The method achieves millisecond-scale inference on consumer hardware and scales to 1000 grooms in 0.3 seconds, while generalizing across hairstyles from straight to curly and across body shapes. Compared with optimization and supervised baselines like GroomGen, Quaffure offers competitive or better hair preservation with significantly faster performance and no data-generation requirements. This enables interactive avatars with complex hairstyles in games and telepresence pipelines.

Abstract

Realistic hair motion is crucial for high-quality avatars, but it is often limited by the computational resources available for real-time applications. To address this challenge, we propose a novel neural approach to predict physically plausible hair deformations that generalizes to various body poses, shapes, and hairstyles. Our model is trained using a self-supervised loss, eliminating the need for expensive data generation and storage. We demonstrate our method's effectiveness through numerous results across a wide range of pose and shape variations, showcasing its robust generalization capabilities and temporally smooth results. Our approach is highly suitable for real-time applications with an inference time of only a few milliseconds on consumer hardware and its ability to scale to predicting the drape of 1000 grooms in 0.3 seconds. Please see our project page here following https://tuurstuyck.github.io/quaffure/quaffure.html

Figures (11)

• Figure 1: We present Quaffure, a real-time quasi-static neural hair simulator, which produces naturally draped hair in only a few milliseconds on commodity hardware, taking the hairstyle, body shape and pose into account. Our method scales to predicting the drape of 1000 hair grooms in just 0.3 seconds. Quaffure is trained using a physics-based self-supervised loss, eliminating the need for simulated training data that is costly and cumbersome to obtain. We show that our method works for a wide variety of body shapes and poses with a range of hairstyles varying from straight to curly, short to long.
• Figure 2: Quaffure Overview: Our method takes a code as input, consisting of a latent code for the rest hair shape, body shape parameters, and full skeleton pose. The output is naturally draped hair produced as the sum of posed hair given the body pose and shape parameters, combined with learned corrections which are produced by the groom deformation decoder. We train our method in two stages: i) an autoencoder is trained on all hairstyles to obtain a groom latent code, and ii) the groom deformation decoder is trained in a physics-based self-supervised fashion. The hair strands are encoded in a 2D texture representation (left) where strands are encoded in the pixel in which the root particle is located. The figure shows how the 3D scalp geometry (bottom) is mapped to a high dimensional 2D texture map (top).
• Figure 3: Pose & Shape-based Deformations: Example of posed groom (top) which accounts for rigid rotations and translations of the rest shape only and the combined posed groom with learned deformations (bottom), which accounts for physical effects such as strand material model, gravity and collisions. Despite body intersections incurred by the rigid transformation applied by the groom transformation module, our proposed network resolves all collisions with the body after applying the learned deformations (bottom). Note the significant change resulting in a natural drape.
• Figure 4: Modeling Diverse Hairstyles: We showcase the versatility of our method in modeling the quasi-static behavior of a plethora of hairstyles and lengths for different body shapes and poses where all results are obtained using the same settings without any manual parameter tuning. Our results cover short/medium/long hair, with various levels of curliness, including straight, wavy, curly, and kinky.
• Figure 5: Resolving Collisions: Our method is conditioned on the body shape parameters, enabling it to efficiently resolve collisions with body shape variations. Here we show a static pose and groom under quasi-statically draped under varying body shapes.