Conditional simulation of complex geological structures using multiple-point statistics (MPS) is presented. Traditional two-point variogram-based methods are insufficient for modeling curvilinear geological features, such as sand channels in clastic reservoirs, which require MPS involving three or more points. Training images, depicting expected geological patterns, are used to infer MPS, which are then incorporated into geostatistical models. This approach allows for the simulation of complex structures by anchoring the inferred statistics to actual data in a sequential simulation mode. The method is tested on a fluvial hydrocarbon reservoir with meandering channels, showing it is simple, general, and efficient for large 3D grids. Keywords: geostatistics, stochastic simulation, training image, random geometry. Traditional simulation methods based on two-point statistics are limited in modeling curvilinear geometries, such as sinuous channels or incised valleys. These require parametrization of specific shapes or joint categorical variability at multiple points. Object-based approaches, while effective for crisp geometries, have limitations in parametrization and conditioning to local data. Pixel-based methods, like simulated annealing and Markov Chain Monte Carlo (MCMC), use MPS but face challenges in computational efficiency and model assumptions. Iterative simulation algorithms based on Gibbs samplers are also proposed. The proposed method uses training images to infer MPS, which are then applied in a sequential simulation to generate realistic geological structures. This approach is efficient, general, and suitable for large-scale 3D simulations.Conditional simulation of complex geological structures using multiple-point statistics (MPS) is presented. Traditional two-point variogram-based methods are insufficient for modeling curvilinear geological features, such as sand channels in clastic reservoirs, which require MPS involving three or more points. Training images, depicting expected geological patterns, are used to infer MPS, which are then incorporated into geostatistical models. This approach allows for the simulation of complex structures by anchoring the inferred statistics to actual data in a sequential simulation mode. The method is tested on a fluvial hydrocarbon reservoir with meandering channels, showing it is simple, general, and efficient for large 3D grids. Keywords: geostatistics, stochastic simulation, training image, random geometry. Traditional simulation methods based on two-point statistics are limited in modeling curvilinear geometries, such as sinuous channels or incised valleys. These require parametrization of specific shapes or joint categorical variability at multiple points. Object-based approaches, while effective for crisp geometries, have limitations in parametrization and conditioning to local data. Pixel-based methods, like simulated annealing and Markov Chain Monte Carlo (MCMC), use MPS but face challenges in computational efficiency and model assumptions. Iterative simulation algorithms based on Gibbs samplers are also proposed. The proposed method uses training images to infer MPS, which are then applied in a sequential simulation to generate realistic geological structures. This approach is efficient, general, and suitable for large-scale 3D simulations.