By-example Synthesis of Compactly Stored Textures

This paper proposes a novel by-example texture synthesis algorithm. Unlike most present algorithms whose synthesized textures of different sizes must be stored in memory until rendering, our results are compactly stored as paths of the generative graphs. This is achieved by adapting the graph-based synthesis framework to be suitable for arbitrary types of textures. Two synthesis processes are carried out successively to grow the texture strip by strip to the desired horizontal and vertical dimensions, each of which is cast as a constrained shortest path problem that can be solved efficiently by our proposed algorithm. The strips are formed by consecutive cuts which are precomputed in the preprocessing step, and are selected with a redesigned mechanism accounting for both global and local matching errors. The variances make the graph-based synthesis framework also applicable to a wide variety of textures besides architectural textures, and the synthesis quality for them is greatly improved.