DGH: Dynamic Gaussian Hair

AI summary

Overview

Research area: Computer vision and computer graphics, specifically digital human modeling — dynamic hair deformation and photorealistic hair rendering.

Technical level: Advanced. The paper assumes familiarity with 3D Gaussian Splatting, neural radiance fields, signed distance fields, volumetric feature encoding, cross-attention, and physics-based hair simulation (XPBD).

Scope: One sentence — the paper presents DGH, a fully data-driven two-stage framework that learns head-motion-driven hair deformation and optimizes a strand-aligned 3D Gaussian hair appearance model for photorealistic dynamic novel-view synthesis.

What This Paper Is About

Creating photorealistic hair that moves realistically is hard because hair involves complex motion, self-occlusion between strands, hair-to-body collisions, and anisotropic light scattering. Existing approaches either rely on physics-based simulation that requires per-hairstyle manual parameter tuning and heavy computation, or reconstruct only static hair that rigidly follows the head. DGH's goal is to learn hair dynamics and appearance directly from data, so that arbitrary hairstyles (long hair, curly hair, ponytails) can be animated by head motion and rendered from novel views without a simulation engine or manual tuning.

Key Contributions

1. Learning-based volumetric hair deformation. A pose-driven, volumetric implicit deformation model that maps static (canonical) hair to deformed hair across diverse hairstyles and strand densities. It is fully data-driven, differentiable, needs no manual parameter tuning, and generalizes to novel head poses.

2. Coarse-to-fine hair deformation dynamics. A two-stage learning framework: a time-independent coarse stage that predicts pose-dependent deformation for stable initialization, and a fine stage that refines high-frequency dynamics via 3D flow prediction with latent-space cross-attention over previous frames.

3. Differentiable dynamic hair appearance optimization. Hair is represented as cylindrical Gaussian primitives (one per hair segment), and a lightweight strand-guided MLP adjusts color (as spherical harmonic coefficients), scale, and opacity using hair tangent and view direction, enabling view-consistent rendering under motion and occlusion.

4. A synthetic dynamic-hair dataset. Because accurate real captures of dynamic hair deformation with strand tracking do not exist, the authors build a dataset with XPBD-based physics simulation and multi-view rendering, which they state they will release.

Main Findings

• Rendering quality beats baselines. On five synthetic hair subjects, the average PSNR/SSIM/LPIPS is 27.009 / 0.871 / 0.127 for DGH, versus 23.073 / 0.847 / 0.131 for Gaussian Haircut and 20.908 / 0.822 / 0.164 for 3DGS. DGH achieves the highest PSNR and SSIM across all subjects; LPIPS is not uniformly best (for example, on Subject 5 Gaussian Haircut reports 0.058 versus DGH's 0.069, and on Subject 1 both report 0.101).

• Deformation is much closer to ground truth than rigid motion. Average L2 deformation error is 0.0832 for DGH versus 0.1639 for rigidly transformed canonical ("rigged") hair; average Chamfer distance is 0.0266 versus 0.0424. Rigid transformation produces unnatural results such as missing gravity effects and hair penetrating the body mesh.

• The SDF constraint prevents penetration. Removing the hair-to-body SDF loss raises the L2 error from 0.0832 (full model) to 0.1269.

• Temporal dynamics and attention both help. Removing the motion (fine) stage raises L2 error to 0.0964; removing cross-attention raises it to 0.0909. The paper notes the coarse stage alone performs similarly to rigid hair on dynamics, and that the fine model plus attention improves motion effects and temporal consistency (reported as flow error ×10⁻³ per timestep).

• Tangent features and curvature-based blending improve appearance. The full appearance model reaches PSNR 28.12 / SSIM 0.90 / LPIPS 0.19. Removing curvature-based blending drops this to 25.08 / 0.88 / 0.18, and removing both tangent features and blending drops it to 20.89 / 0.80 / 0.19. The authors attribute gains to tangent conditioning sharpening local detail and curvature blending smoothing transitions between adjacent Gaussians in high-curvature regions.

• Integration with Gaussians is straightforward. DGH can be merged with body Gaussian primitives for avatar re-animation, since it needs only head rotation and a dense point cloud or pre-trained Gaussian Splatting avatar, with no rigging or simulation overhead.

Methodology in Plain English

The pipeline has two stages, both trained on a single A100 GPU with the Adam optimizer and a learning rate of 1×10⁻⁴.

Stage I — learning how hair moves. The static hair point cloud is rigidly transformed by the head pose, then voxelized into a signed distance field volume; a proxy mesh of the posed head and shoulders is voxelized the same way. 3D CNNs encode both volumes into feature grids. For each sampled hair point, a feature is interpolated from this grid and fed, along with positional encoding and the head pose, into an MLP that predicts a displacement. This coarse step is time-independent and trained with an MSE point loss plus an SDF penalty that pushes points out of the body mesh. The fine step then voxelizes the two previous hair states, encodes them, and uses cross-attention to build a fused flow volume. Each point from the previous frame samples this volume, concatenates the current pose, the previous flow, and positional encoding, and a refinement network predicts a per-point 3D flow vector supervised against ground-truth flow. Inference is recurrent: at t=0 only the coarse stage runs, at t=1 the fine stage runs with self-attention on t=0 and zero input flow, and for t>1 the pipeline proceeds recurrently.

Stage II — learning how hair looks. Each hair strand is represented as a chain of connected cylindrical Gaussians whose length far exceeds their radius, which removes the need for Gaussian densification and keeps the primitive count fixed across frames. Because the deformation model can warp Gaussians from the canonical frame to any frame, color and scale propagate trivially over time — but doing this naively loses appearance detail due to self-occlusion and light-hair interaction. So a lightweight MLP takes features sampled from the encoded deformed hair volume plus positional encodings of position, hair tangent, and view direction, and outputs modified spherical harmonic color, scale, and opacity. A curvature-based blending step computes normalized local curvature as a weight and uses it to blend spherical harmonics and opacity between neighboring segments, reducing shading discontinuities where adjacent tangents differ sharply. The model is trained with L1, SSIM, and LPIPS losses (weights 1.0, 0.1, 0.1; Stage I weights are 1.0 for point loss and 0.01 for SDF).

Data. The synthetic dataset uses XPBD-based physics simulation. Each simulated groom consists of 1500k hair strands with 24 vertices per strand. For each hairstyle, 100 motion sequences of 100 frames are simulated, totaling 10k frames, each including deformed hair positions, head motion parameters, and the upper-body mesh. For appearance, 500-frame sequences per groom are rendered as multi-view videos from 24 camera angles in Blender. Evaluation covers 5 hair subjects, with 90 motion sequences used for training and the remaining 10 for testing per groom; dynamic appearance is assessed on unseen motions (100 frames per subject) and novel views captured via horizontal camera rotation (100 views).

Why This Matters

The work points to a practical alternative to physics-based hair simulation for animatable avatars: hair dynamics become a learned, differentiable function of head motion rather than a hand-tuned simulation with expensive collision handling. For research, it connects dynamic hair with the 3D Gaussian Splatting avatar line of work, which until now typically treated hair as a static transformation rigidly attached to the head. It also contributes a synthetic dynamic-hair dataset with frame-by-frame geometry and multi-view renders, addressing the stated lack of real dynamic-hair captures with strand tracking.

Real-world applications:

• AR/VR avatars and telepresence — animatable, photorealistic hair driven by head motion for virtual characters.
• Animation and VFX pipelines — reducing manual simulation parameter tuning per hairstyle and costly high-end rendering for dynamic hair.
• Game and real-time engines — a learned deformation and appearance model that avoids the full cost of path-traced hair rendering, which the paper notes scales to a render farm with multiple GPUs.
• Digital humans and virtual try-on / content creation — hairstyle-agnostic deformation that generalizes across long, curly, and ponytail grooms and integrates with existing Gaussian avatar systems.

Industry relevance: the author affiliations include Meta Reality Labs Research alongside the University of Southern California, and the paper explicitly frames its motivation around AR/VR device constraints — where physics-based simulators are described as impractical due to limited GPU memory and processing power.

Future Directions

• Closing the realism gap with simulation. The physics-based XPBD results are treated as an upper bound rather than a learned target, leaving open how close data-driven dynamics can get to simulation quality on hair behaviors the training distribution does not cover.
• Generalizing beyond the synthetic dataset. All evaluation uses the paper's own synthetic data; validating on real captured hair with accurate strand tracking remains unaddressed given the stated lack of such data.
• Handling more hair types and environmental interactions. The paper reports results on long, curly, and ponytail styles with upper-body collision; broader interactions such as wind, hands, and other objects are not evaluated.
• Efficiency of recurrent inference. Because the fine stage runs recurrently and depends on the previous two frames, questions remain about error accumulation over long sequences and runtime/memory behavior — details the authors defer to the Appendices.

Target Audience

Researchers and practitioners in computer vision, computer graphics, and digital human modeling — particularly those working on 3D Gaussian Splatting avatars, neural hair reconstruction and animation, and AR/VR avatar systems. It will also interest technical artists and rendering engineers looking for a learned alternative to physics-based hair simulation, and readers who need a reference synthetic dataset for dynamic hair deformation and multi-view appearance. Beginners will find the method sections dense, since the paper assumes working knowledge of SDF volumes, volumetric feature grids, cross-attention, and Gaussian rasterization.

Authors’ 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.