Paweł Skierś

Deep generative models, and in particular, diffusion-based approaches, are now able to deliver an unprecedented image and video synthesis quality. However, despite the high fidelity of the samples, the distribution learned by these methods often significantly deviates from the real data distribution. This failure of the model to properly generalize to the true data distribution is known to manifest itself in several ways, including hallucinations, mode collapse, or, in the most drastic cases, memorization of the training data. While these problems are not necessarily detrimental when the main goal is visual realism, they pose serious challenges whenever precise distribution modeling is required, for instance, in scientific simulations or medical imaging. The goal of this doctorate is to deepen theoretical understanding of generalization in the diffusion model and provide practical insights, allowing for more accurate modelling of the data distribution.