Abstract : We present the concept of hierarchical instantiation for radiosity. This new method enables an efficient, yet accurate determination of the illumination in very large scenes, where similar objects are replaced by instances of the same element. Instances are equipped with suitable radiative properties and are used to replace large amounts of geometry at multiple levels of the scene hierarchy. In essence, our algorithm replaces a single very large hierarchical radiosity problem by a collection of hierarchical radiosity problems within small sets of objects at a time, at several hierarchical levels. We prove the applicability of our method on architectural scenes with replicated geometry. However we reach the best time and memory gains on plant models thanks to the high degree of self-similarity in such kinds of scenes. This allows us to compute lighting simulations on scenes including a very large number of polygons in a short time on machines with limited memory.