The computational cost of predicting wildland fire spread across large, diverse landscapes is significant using current models, which limits the ability to use simulations to develop mitigation strategies or perform forecasting. This paper presents a machine learning approach to estimate the time-resolved spatial evolution of a wildland fire front using a deep convolutional inverse graphics network (DCIGN). The DCIGN was trained and tested for wildland fire spread across simple homogeneous landscapes as well as heterogeneous landscapes having complex terrain. Data sets for training, validation, and testing were created using computational models. The model for homogeneous landscapes was based on a rate of spread from the model of Rothermel, while heterogeneous spread was modeled using FARSITE. Over 10,000 model predictions were made to determine burn maps in 6 h increments up to 24 h after ignition. Overall the predicted burn maps from the DCIGN-based approach agreed with simulation results, with mean precision, sensitivity, F-measure, and Chan–Vese similarity of 0.97, 0.92, 0.93, and 0.93, respectively. Noise in the input parameters was found to not significantly impact the DCIGN-based predictions. The computational cost of the method was found to be significantly better than the computational model for heterogeneous spatial conditions where a reduction in simulation time of 10^2 −10^5 was observed. In addition, the DCIGN-based approach was shown to be capable of predicting burn maps further in the future by recursively using previous predictions as inputs to the DCIGN. The machine learning DCIGN approach was able to provide fire spread predictions at a computational cost three orders of magnitude less than current models.