Sarah Jobalia

How we encode virtual 3D data, especially in cases of generative AI, can either restrict or facilitate diversity and realism. Ideal encodings capture the social richness of our experience as well as the physical realities of our world. These representations are particularly important in the case of virtual hair, which is both computationally complex and culturally important to represent faithfully.

Disentangling strand texture from style highlights two fundamental but separate aspects of hair, allowing groom construction that can capture the essence of desired styles. Drawing upon scientific, artistic, and cultural expertise, we define strand texture as the various distinctive patterns formed by forces internal to a hair strand. We propose a novel taxonomy that encompasses naturally occurring hair strand textures across all hair types.

From this taxonomy, we build a parametric encoding that is both qualitatively accessible, allowing users to readily locate their own hair within the parameter space, and quantitatively precise, allowing the generation of individual strands from texture inputs. We demonstrate the effectiveness of this encoding through a diffusion model that can generate strands matching an input texture with minimal training data. With the aid of this model, we also build a texture-transfer pipeline that can change the hair texture of any input hairstyle, further demonstrating the benefits of disentanglement.