||The will for realism in computer generated pictures has led to an increase in the number of details modelised. Actual production pipelines handle models of several hundreds of billions of polygons and several Go of textures, all those objects made to support close inspection with pictures of thousands of pixels. Actual rendering methods are thus showing their limitations to render in real time such models, every step of the rendering pipeline resting on over sampling the details. Most solutions plan to reduce the number of samples required to produce a quality picture but as they are data driven, they are not scalable. Another approach is to build model based representations mixing geometry and appearances of the objects to unify modelisation and rendering, and thus avoid the need to over sample the data. This is this kind of solution that are being developed actually in R3AM; building a pre filtered representation of the shape and appearance of an object or of a group of objects. The first attempts are for now limited to surfaces but they give very good results. The main point of this PhD in to follow this path to improve and generalize those models.