Animating numerous humans in virtual worlds with real-time constraint is difficult. The problem is mainly due to the computation time required to animate complex motions. This problem can be solved by the use of animation levels of detail. These manage the computation complexity by selecting the way each model is computed. An animation level of detail of an object consist in a set of animation models with different computation costs. In this paper, we propose a new approach of model design in order to easily build animation levels of detail. Our model architecture consists in a pipeline of sub-models. Each sub-model performs a given task in the animation process: animation of the body or of the legs for example. It encapsulates several computation methods for the given task with different computation costs. Our animation levels of detail minimize the complexity by selecting the best suitable animation method according to the environment. The policy of model selection also depends on external parameters: distance to the point of view and visibility. We apply our approach to the building of a walking model on complex grounds.