Massively parallel layout generation in real time

Abstract

Conceiving an artwork requires designers to create assets and organize (or layout) them in a harmonious, self-orating story. While creativity is fundamental to both aspects, the latter can be bolstered with automated techniques. We present a first true SIMD formulation for the layout generation and leverage CUDA-enabled GPU to scan through millions of possible permutations and rank them on aesthetic appeal using weighted parameters such as symmetry, alignment, density, size balance, etc. The entire process happens in real-time using a GPU-accelerated implementation of replica exchange Monte Carlo Markov Chain method. The exploration of design space is rapidly narrowed by performing distant jumps from poorly ranked layouts, and fine tuning the highly ranked ones. Several iterations are carried out until desired rank or system convergence is achieved. In contrast to existing approaches, our technique generates aesthetically better layouts and runs more than two orders of magnitude faster.