DGH: Dynamic Gaussian Hair

TL;DR

Dynamic Gaussian Hair (DGH) introduces a two-stage, data-driven framework that learns both dynamic hair deformation and appearance within a differentiable Gaussian representation. By combining a coarse-to-fine dynamic hair model with a strand-guided Gaussian appearance network, DGH achieves photorealistic, view-consistent hair rendering under motion and integrates with Gaussian avatar systems. The approach is validated on a synthetic dataset, with ablations confirming the contributions of collision constraints, motion refinement, and curvature-based appearance blending. DGH offers a scalable, data-driven alternative to physics-based hair simulation and paves the way for animatable, high-fidelity Gaussian avatars in virtual production and AR/VR applications.

Abstract

The creation of photorealistic dynamic hair remains a major challenge in digital human modeling because of the complex motions, occlusions, and light scattering. Existing methods often resort to static capture and physics-based models that do not scale as they require manual parameter fine-tuning to handle the diversity of hairstyles and motions, and heavy computation to obtain high-quality appearance. In this paper, we present Dynamic Gaussian Hair (DGH), a novel framework that efficiently learns hair dynamics and appearance. We propose: (1) a coarse-to-fine model that learns temporally coherent hair motion dynamics across diverse hairstyles; (2) a strand-guided optimization module that learns a dynamic 3D Gaussian representation for hair appearance with support for differentiable rendering, enabling gradient-based learning of view-consistent appearance under motion. Unlike prior simulation-based pipelines, our approach is fully data-driven, scales with training data, and generalizes across various hairstyles and head motion sequences. Additionally, DGH can be seamlessly integrated into a 3D Gaussian avatar framework, enabling realistic, animatable hair for high-fidelity avatar representation. DGH achieves promising geometry and appearance results, providing a scalable, data-driven alternative to physics-based simulation and rendering.

Figures (24)

• Figure 1: Dynamic Gaussian Hair (DGH) is a framework that learns dynamic deformation and photorealistic novel-view synthesis of arbitrary hairstyles driven by head motions, while respecting upper-body collision. At runtime, given a hairstyle and head motion (a), DGH infers initial hair deformations (b), refines the deformations with dynamics (c), and generates 3D Gaussian Splats to achieve photorealistic novel-view synthesis (d).
• Figure 2: Framework Overview. DGH learns hair deformation dynamics and photorealistic appearance. Stage I: Coarse-to-Fine Dynamic Hair Modeling. The input hair model and the upper body are transformed into a canonical hair volume $V^{\text{rigid}}_{\text{hair}}$ and a pose volume $V_{\text{pose}}$, respectively. Then, a coarse-to-fine strategy deforms the hair model. At the coarse stage, points $\mathbf{p}_i$ are sampled from the rigidly transformed hair, and the interpolated features from $\mathcal{E}_{\text{pose}}$, $\mathcal{E}_{\text{hair}}$, head pose $\mathcal{H}$, and positional encoding $E(p)$ are concatenated and fed into an MLP $\mathcal{M}$ to predict displacements $\Delta \mathbf{p}$, producing deformed hair points ${P}_{\text{hair}}$. The fine stage refines hair deformation with dynamics by estimating flow $\mathcal{F}_{\text{flow}}^t$ through cross-attention between volumetric features from previous frames $V_{\text{hair}}^{t{-}2}$ and $V_{\text{hair}}^{t{-}1}$, ensuring smooth temporal transitions. Stage II: Appearance Optimization. We train an MLP $\mathcal{D}$ to predict color $c'$, scale $s'$, and opacity $\alpha'$ of 3D Gaussian Splats from features of the deformed hair. Differentiable rasterization leverages the appearance model to synthesize high-quality renderings that adapt to hair movement and occlusion dynamics.
• Figure 3: Hair Representation. We show different hair representations with tangent vectors and curvature. For Gaussian hair, we attach cylindrical Gaussian primitive luo2024gaussianhair to each segment with a length much greater than its radius.
• Figure 4: Hair Deformation Comparison. From left to right column, given a canonical groom and upper body pose, rigid transformations result in unrealistic hair deformation with upper-body penetration, while our method achieves natural deformation across different grooms with correct collisions.
• Figure 5: Dynamic hair appearance evaluation. The left section compares dynamic hair rendering using our hair deformation/tracking model against baseline methods: 3DGS kerbl20233d and Gaussian Haircut (GH) zakharov2024human. The right section shows rendering results merging hair and body Gaussian primitives. Without our hair tracking model, the hair appears rigid and unrealistic, while our appearance model enhances hair rendering quality.