The work presented in this paper aims to use level-of-detail representation in realizing interactive walkthroughs or ignoring useless details in large architectural models. In order to choose the right representation of a model, we have to evaluate the error committed when using a simplified version instead of the full description of an object. This error depends on the object deformation during the simplification process but also on the importance of this object in the current viewing conditions. This “visiblei error is used with different visualization strategies to find the model representation which satisfies either a quality criterion or a cost condition.