Ceramic matrix composites are often the basis for components meant to be used under severe conditions: aircraft brakes, rocket engine hot parts, plasma facing elements of the Tokamak... These materials are mainly obtained by Chemical Vapor Infiltration: matrix is formed by heterogeneous reaction of a gas precursor penetrating the fibrous preform. Quality and properties of the densified part is highly dependent of elaboration parameters such as temperature, pressure or even gas concentration... Experimental optimization of the process proves long and costly and triggered the need for a numerical model. In this perspective, a multi-scale approach of the densification process is considered, based on a 3D representation of the preform. X-Ray tomography images are used as they render a detailed capture of fiber and tow organization and before actual simulation, this virtual preform is modeled by 3D image analysis. The numerical densification tool consists of two algorithms: one for pore-scale simulation, the other for macroscopic infiltration. In both cases, gas evolution is recreated by a Pearson's random walk technique whilst chemical reaction is represented by a discrete VOF method. The multi-scale method consists in determining micro scale diffusion and reaction laws that will be used as input for configuration of the macroscopic model. Our overall methodology and the segmentation techniques will be presented. Results of a densification simulation will then be depicted and discussed.