Image-based Fitting of Procedural Yarn Models

Abstract

While common in real life, rendering fiber and cloth accurately is challenging. Recent fiber-based, procedural rendering approaches proved to be able to capture a great amount of details of real yarn. However, the current automatic method of fitting the model parameters is expensive and inaccessible as it relies on micro CT scans of the reference yarn. The alternative is to have an artist fit the parameters by hand, which is impractical because of the large number of parameters. We present a proof-of-concept for a purely image-based approach to fit the parameters of a procedural yarn model. Using gradient descent and pixel-based loss functions, we are able to extract a subset of the model parameters from rendered images with known parameters. The appearance of the fitted models is nearly indistinguishable from the reference images.