Join out talk on Thursday 14/08 in the "Packing In Structure" session at 3:45pm (West Building, Rooms 211–214) to see this in action!

The gradient-free optimization enables fast optimization over both continuous and discrete parameters, while GPU-accelerated SDF evaluation keeps the optimization loop interactive. Together, they let us achieve real-time interaction even for complex, highly parameterized procedural patterns.

Each new sketch triggers a gradient-free optimization process that minimizes a weighted per-pixel distance between the pattern’s SDF and the stroke SDF. The weights extracted from the both pattern and "Void" strokes enable focusing on drawn regions as well as empty ones.

We model a procedural pattern generator as a sequence of functions called Levels, where each Level’s parameters form a superset of the previous one
This enables coarse-to-fine optimization, where earlier parameters are inherited and new ones are optimized.