Towards practical rendering of fiber-level cloth appearance models

Accurate representation of realistic cloth appearance is of high importance in many industry fields such as entertainment and textile design. However, microstructure of fibers and their optical properties generate very complex lighting effects, often not reproduced by empirical and theoretical models. In contrast, data-driven appearance models obtained with simulations on explicit representations of fibers have proved to yield accurate cloth appearance, but resulting in discretized distribution functions that require costly precomputations, massive storage, and expensive evaluation in render time. Finding efficient representations for these models is therefore of key importance to find good trade-offs between accuracy and affordable computational costs. In this work we explore the use of different analytical models to represent these data-driven distributions, which arises as a promising middle-ground solution to this problem with benefits in both storage, computational cost, and affordable generation of new fiber appearance models. We base this analysis on our recent work where we provide highly detailed tabulations of different types of cloth fibers appearance. We analyze the spectral component of different fiber appearance functions, and observe that just ten levels of spherical harmonics are sufficient to represent the appearance many smooth fibers. We also propose a generic method to fit Gaussian Mixture Models to massively tabulated appearance functions, reducing storage costs from hundreds of MBs to a few KBs, and producing equivalent results 40 times faster. We finally analyze how interpolations in the space of fibers absorption can be exploited to generate novel fiber appearance functions without requiring costly brute-force precomputations.