Parallel Controllable
Texture Synthesis
Abstract
We present a texture synthesis scheme based on neighborhood matching,
with contributions in two areas: parallelism and control. Our scheme
defines an infinite, deterministic, aperiodic texture, from which
windows can be computed in real-time on a GPU. We attain high-quality
synthesis using a new analysis structure called the Gaussian stack,
together with a coordinate upsampling step and a subpass correction
approach. Texture
variation is achieved by multiresolution jittering of exemplar
coordinates. Combined with the local support of parallel synthesis, the
jitter enables intuitive user controls including multiscale randomness,
spatial modulation over both exemplar and output, feature drag-and-drop,
and periodicity constraints. We also introduce synthesis magnification,
a fast method for amplifying coarse synthesis results to higher
resolution.
Links
|
