Real-Time Virtual Character Production Pipeline: Hair, Skin, and Physics Rendering Techniques

This article introduces the HuànXing Digital Human platform, a 3D digital‑human solution jointly developed by the Industrial Production Department and the AI Platform Department. The platform consists of an industrialized asset generation pipeline and an AI‑driven facial‑and‑motion capture system, supporting various use cases such as live streaming, virtual‑content creation, user avatars, virtual gifts, and virtual‑idol IPs.

The technical focus is on the asset generation pipeline, covering realistic hair rendering, personalized character customization, and physics simulation. The hair rendering section compares two mainstream approaches: Hair‑Card/Mesh (geometry‑based, low‑cost, suitable for mobile and low‑end hardware) and Strand‑Based (physically accurate, high‑quality, suitable for virtual idols). A detailed table contrasts modeling difficulty, lighting quality, and performance of the two methods.

For Strand‑Based hair, the pipeline implements single scattering (using the Marschner model), multiple scattering (via local and global scattering with Deep Opacity Maps), and visibility calculation through a Visibility‑Buffer that stores per‑pixel strand information. Hair‑Mesh visibility is handled with alpha‑test transparency, dithering, alpha‑to‑coverage, and temporal super‑resolution, while semi‑transparent rendering uses layered rendering with depth‑based sorting.

The skin rendering part addresses the challenge of realistic subsurface scattering (SSS). The platform adopts a Subsurface Profile (SSP) model with diffusion profiles, using multi‑Gaussian fitting (up to six Gaussians) to approximate the scattering curve. Different real‑time SSS techniques are discussed: the 4‑pass separable SSS (4s), Burley Normalized Diffusion, and a pre‑integrated lookup‑table approach for low‑end devices. Each method balances visual fidelity and performance.

Eye rendering includes modeling of sclera, limbus, iris, pupil, and cornea, with techniques such as matcap‑based eye‑glint, refraction via IOR or parallax mapping, and multi‑layer scattering using depth‑based texture channels.

Character personalization is achieved through a bone‑driven facial rig (high‑degree‑of‑freedom, low‑cost) complemented by blendshape‑based expression animation. A dedicated facial‑configuration tool allows designers to map facial controls to specific bones and ranges, reducing iteration time. The system also supports makeup, accessories, and clothing changes.

The clothing system uses skeletal meshes with GameplayTags to define hierarchical body parts and clothing slots, enabling flexible mix‑and‑match outfits and material swaps.

Physics simulation covers hair dynamics (DynamicBone for low‑cost, strand‑based simulation for high‑quality) and cloth simulation using Position‑Based Dynamics (PBD) with stretch and bend constraints. The article outlines the limitations of PBD and mentions future adoption of Extended Position‑Based Dynamics (XPBD) via UE5’s Chaos system.

Finally, the roadmap envisions AI‑assisted animation generation, AI‑driven concept art, and continued optimization of the pipeline to provide every B‑Station user with a personalized 3D avatar.