Whereas outcrop models can provide important information on reservoir architecture and heterogeneity, it is not entirely clear how such information can be used exhaustively in geostatistical reservoir modeling. Traditional, variogram-based geostatistics is inadequate in that regard because the variogram is too limiting in capturing geologic heterogeneity from outcrops. A new field, termed multiple-point geostatistics, does not rely on variogram models. Instead, it allows capturing structure from so-called training images. Multiple-point geostatistics borrows multiple-point patterns from the training image, then anchors them to subsurface well-log, seismic, and production data. However, multiple-point geostatistics does not escape from the same principles as traditional variogram-based geostatistics: it is still a stochastic method and, hence, relies on the commonly forgotten principles of stationarity and ergodicity. These principles dictate that the training image used in multiple-point geostatistics cannot be chosen arbitrarily and that not all outcrops might be suitable training image models. In this paper, we outline the guiding principles of using analog models in multiple-point geostatistics and show that simple, so-called modular training images can be used to build complex reservoir models using geostatistics algorithms.