In contrast to the conventional view that facies distribution patterns on carbonate ramps are relatively simple, outcrop analogue studies point to a high degree of internal facies complexity. Depending on the diagenetic overprint, this complex pattern may result in reservoir compartmentalization due to the presence of interflow baffles. The often subseismic scale heterogeneities may not be included in conventional reservoir modelling. In order to evaluate how facies heterogeneities in shoal reservoirs can be modelled realistically, this paper presents a facies modelling workflow which includes a new approach to the design of training images for multiple-point statistics (MPS). The workflow was developed in the course of a reservoir outcrop analogue study of a Ladinian (Middle Triassic) coquinadominated shoal complex in SW Germany which was deposited on an epicontinental, gently inclined carbonate ramp. The data set was based on an intensive field study and includes 3D facies and sequence stratigraphic analyses of the largest shoal complex in the Quaderkalk Formation (Upper Muschelkalk). This several metre thick shoal complex represents a subseismic scale, bioclast-rich reservoir analogue and has a very heterogeneous facies pattern. Integrating 1D facies logs and sequence stratigraphic trends from tens of outcrop sections and cores, two nested 3D geocellular facies models were produced: (i) a large-scale (30 x 30 km) model based on truncated Gaussian simulation (TGS); this formed the basis for (ii) a smaller-scale (10 x 10 km), more detailed model based on multiple point statistics. In addition, a new approach for training image design was developed to honour small-scale sequence stratigraphic trends and lateral facies patterns observed in modern analogues. Compared to facies patterns in modern analogues, the large-scale model presents geologically-feasible facies distribution patterns and geometries, and in addition shows a vertical facies distribution which is similar to the observed sequence stratigraphic architecture of the outcrop data-set used. Due to the new training image design, the final small-scale model has a distribution pattern of facies heterogeneities which looks similar to modern facies distributions in the offshore UAE and thus represents a valuable method of producing realistic reservoir facies models. The modelling workflow and the new approach for training image design presented will help to reduce uncertainties in the understanding and modelling of subsurface reservoirs by using a systematic combination of outcrop data and modern analogues, with the consistent application of sequence stratigraphic principles. In addition, this study emphasises the importance of careful training image design, derived from modern analogues, which can be used as realistic inputs in order to optimize multiple point simulations, and which may be applied to producing bioclastic reservoirs such as those located on the Arabian Plate or offshore Brazil.