We present the first results of a research aimed at generating useradapted image descriptions from annotated knowledge sources. This system employs a User Model and several knowledge sources to select the image attributes to include in the description and the level of detail. Both ‘individual’ and ‘comparative-descriptions’ may be generated, by taking an appropriate ‘reference’ image according to the context and to an ontology of concepts in the domain to which the image refers; the comparison strategy is suited to the User background and to the interaction history. All data employed in the generation of these descriptions (the image, the discourse) are annotated by a XML-like language. Results obtained in the description of radiological images are presented, and the advantage of annotating knowledge sources are discussed.