Section: New Results

Texture Synthesis

Understanding and controlling contrast oscillations in stochastic texture algorithms using Spectrum of Variance

Participants : Fabrice Neyret, Eric Heitz.

We identify and analyze a major issue pertaining to all power-spectrum based texture synthesis algorithms from Fourier synthesis to procedural noise algori thms like Perlin or Gabor noise, namely, the oscillation of contrast (see Figure 13). One of our key contributions is to introduce a simple yet powerf ul descriptor of signals, the Spectrum of Variance (not to be confused with the PSD), which, to our surprise, has never been leveraged before. In this new framework, several issues get easy to understand measure and control, with new handles, as we illustrate. We finally show that fixing oscillation of contra st opens many doors to a more controllable authoring of stochastic texturing. We explore some of the new reachable possibilities such as constrained noise content and bridges towards very different families of look such as cellular patterns, points-like distributions or reaction-diffusion [17].

Figure 13. Power-spectrum based texturing algorithms (e.g., Gabor, Fourier synthesis) suffer from unexpected low frequency contrast variations (a,b,c top) even when the spectrum has no low frequency (the contrast field is display in red in (c)). This prevents precise authoring with non- linear transform, like color LUT (b top). Our renormalization method allows to control the stationarity (a,b,c bottom). It also opens many doors for noise authoring such as the generation of reaction-diffusion-like strips and spots (b bottom), cellular-like patterns (d), content constraints (e), or the parame trization of height maps relative to local extrema (f).