Abstract: A procedural pattern generation process, called multi-scale "assemblage" is introduced. An assemblage is defined as a multi-scale composition of "multi-variate" statistical figures, that can be kernel functions for defining noise-like texture basis functions, or that can be patterns for defining structured procedural textures. This paper presents two main contributions: 1) a new procedural random point distribution function, that, unlike point jittering, allow us to take into account some spatial dependencies among figures and 2) a "multi-variate" approach that, instead of defining finite sets of constant figures, allows us to generate nearly infinite variations of figures on-the-fly. For both, we use a "statistical shape model", which is a representation of shape variations. Thanks to a direct GPU implementation, assemblage textures can be used to generate new classes of procedural textures for real-time rendering by preserving all characteristics of usual procedural textures, namely: infinity, definition independency (provided the figures are also definition independent) and extreme compactness.