We present a new method of dynamic light effect generation using stochastic models. Similar to dynamic lighting scenes in nature, the resulting light effects are unpredictable, yet recognizable. Furthermore, we present a method to learn the stochastic models from a video source of a natural scene. The method extracts the representative colors from the video and subsequently learns the typical transitions between the colors. After the model has been learned, the rendering of the effects has low memory and processing requirements, making it suitable for implementation even on embedded platforms. The recognition of the produced light effects was tested using a large user base and three automatically created models and a hand crafted one. The results show the suitability of the method for dynamic atmosphere creation, but also a high appreciation of the produced light effects.