Owning sets of models is crucial in many fields, so as to validate concepts or to test algorithms that handle models, model transformations. Since such models are not always available, generators can be used to automatically generate sets of models. Unfortunately, the generated models are very close to each others in term of graph structure, and element naming is poorly diverse. Usually, they badly cover the solutions' space. In this paper, we propose a complete approach to generate meaningful and diverse models. We use probability simulation to tackle the issue of diversity inside one model. Probability distributions are gathered according to domain quality metrics, and using statistical analysis of real data. We propose novel measures to estimate differences between two models and we provide solutions to handle a whole set of models and perform several operations on these models: comparing them, selecting the most diverse and representative ones and graphically observe the diversity. Implementations that are related to difference measurement are gathered in a tool named COMODI. We applied these model comparison measures in order to improve diversity in Model Driven Engineering using genetic algorithms.