Thesis title: Editing Methods for Procedural Materials
Procedural modeling has long offered resolution-independent, editable
representations for computer graphics, yet the inverse estimation of
procedural parameters, interactive editing, and collaborative version control
remain challenging. In this work, we present a suite of complementary techniques
that tackle these issues. First, we introduce a method for recovering the
parameters of procedural vector patterns, defined as collections of standard
graphics primitives arranged by arbitrary programs, by differentiating their
Signed Distance Fields rather than relying solely on raster-based losses.
This approach yields reliable parameter estimation even when target images
are hand-drawn or comprise sparse overlapping elements. Second, we propose an
interactive system to sketch procedural patterns, where complex designs
are decomposed into a hierarchy of coarse-to-fine levels, namely scaffolds. Users
can iteratively refine their sketches while a gradient-free, GPU-accelerated
optimizer updates only the additional parameters introduced at each scaffold
level. Finally, we develop a version control system for procedural node
graphs. By directly comparing typed node structures using a greedy matching
algorithm, which approximates the Graph Edit Distance under type constraints,
our method reliably diffs and merges concurrent edits in collaborative settings.
Extensive experiments and comparisons with state-of-the-art approaches prove
that our work enhances both the quality and efficiency of procedural
content creation workflows.