In this paper, we propose a generative model for representing complex motion, such as wavy river, dancing fire and dangling cloth. Our generative method consists of four components: (1) A photometric model using primal sketch[8] which transfers an image into an attribute graph representation. Each vertex of the graph is a scaled and oriented image patch selected from a dictionary. The graph connects and aligns these patches. (2) A geometric model which characterizes the deformation of the attribute graph. (3) A dynamic model, which specifies the motion dynamics of these vertices (patches) and their interactions in the form of coupled Markov chains. (4) A topological model, which interprets the graph topological changes over time. We learn this generative model by a stochastic gradient algorithm implemented by Markov Chain Monte Carlo (MCMC) sampling. This method is shown to be effective in handling the topological changes of graphs. The correctness of the learned model is verified by the low-dimension reconstruction of the original image as well as by the realistic motion sequences it synthesized.