Diffusion models are major contenders when it comes to generation of images or data. They generate outputs applying a denoising process to random inputs which is learned as reversal of a stochastic process that adds noise to a training sample. But the underlying causal aspect of this process remains unexplored. Is there an inherent causal structure created inside of the generating process? If so, how is this represented inside of the model after training? In January we will be researching this specific area with two research assistants by training a transformer-based diffusion model with the goal to evaluate statistical dependence structures in the reverse process. Our method involves examining the conditional dependencies between the components generated in the latent space of an image. Specifically, we will analyse how different regions of attention within the model are influenced by the drift term. Understanding of how causal structures are formed in generative tasks create a new way to comprehend and possibly evaluate these models.