Summary: Texture authoring is an important part of creating virtual 3D worlds. The constant increase of complexity requires to set up automatic, scalable, controllable, and efficient texture synthesis algorithms. Procedural textures have been successfully used over decades because they provide data compression and genericity. They are defined by sets of parameters and by generative rules. However, the creation of procedural textures is a tedious task requiring expert knowledge and trial and error authoring. Many procedural texture synthesis algorithms rely on a random function called noise. The synthesis of spatial noise functions is thus a core research issue. Since Perlin's seminal work, various procedural noise synthesis algorithms have emerged, all having a different focus and, subsequently, different properties. Some key properties are: computational complexity, locality of computations (leading to optimal parallelism), spectral control, scalability to higher dimensions, mapping on surfaces without parametrization, accurate antialiasing, exact differentiability (continuity and resolution independence), visual quality (no alignment artifacts, no repetitions, no contrast losses, etc.) and pattern diversity (e.g. possible extensions to non-Gaussian processes). In this talk, we present our contributions to the synthesis of noise functions and their use for procedural texture modeling. We present a noise model based on local Fourier series, a bilayered texture model and semi-procedural texture synthesis, where a noise-similar basis function is used to create procedural binary structures, which are then enhanced with color details using example-based texture synthesis.