The objective of the study is to present a method for improving the capability of a semi-physical network model to predict large fire patterns in heterogeneous landscapes. The method, which can be viewed as an Autonomous System, consists in generating an amorphous network by sowing vegetation cells on-the-fly. All the information on fire behavior is contained in the very few digital elevation map pixels close to the fire front, in which fuel items are heated or burning. The method is applied to two distinct scenarios: a no-wind and no-slope academic case and a historical Mediterranean fire that occurred in the South-East of France in 2005. Both cases are discussed in terms of CPU time and memory allocation gains.