Bi‐Layer textures: a Model for Synthesis and Deformation of Composite Textures

Abstract

Our noise model decomposes an input exemplar as a structure layer and a noise layer. The noise layer captures a spatially varying Gaussian noise as a blend of stationary noises. Large outputs are synthesized on-the-fly by synchronized synthesis of the layers. Variety can be achieved at the synthesis stage by deforming the structure layer while preserving fine scale appearance, encoded in the noise layer.

Local spot noise for procedural surface details synthesis

“…Authors used spatial deformations and turbulence on the output to break periodicity. Instead of using phase information to reproduced the input example structure, Guingo et al [25] proposed to separately model the structural component by using spatial methods like patch-based methods and the stochastic component with random phase noises.…”

Procedural phasor noise

“…We believe this is a novel tool in procedural synthesis that will lead to further developments. In particular, procedural noise is typically used to add fine scale unstructured details [Guingo et al 2017]. Phasor noise paves the way to the ability to synthesize increasingly structured patterns from purely procedural synthesizers.…”

Cyclostationary Gaussian noise: theory and synthesis

“…Gilet et al [GSV + 14] shows how fixing some phases can help preserve the spatial organization of an exemplar; however, freezing phases makes the output tend towards a perfectly periodic signal, which comes at the expense of variety. Guingo et al [GSDC17] explores the possibility of using several spectra to provide a higher control over the output spatial organization, but without synthesizing near‐regular textures. Phasor noise [TEZ + 19] generate patterns with strong variations of intensity, by controlling the instantaneous phase and a wave profile.…”

“…They are thus unable to synthesize periodic statistics, and subsequently unable to model color transformations in near‐regular textures. Works attempting to control the spatial organization of the underling stationary process do so by either incidentally altering the stationarity [GSDC17, PGDG16, CGG19], by varying the sparse convolution using specific point processes; or, by altering their Gaussian nature through the control of the phase [GSV + 14, TEZ + 19]; or by reducing the convergence towards a Gaussian process [HN18]. Conversely, we propose to alterate stationarity and keep the Gaussian nature by exploiting a generalization to Gaussian cyclostationary processes.…”