Irregularities of Latent Space Geometry in Diffusion Models

This paper extends experimental evidence of high-Lipschitz irregularities in diffusion mappings from the prior distribution to the data distribution. Previously, it was observed that for diffusion models, slight variation of the initial Gaussian noise results in notably different generated images. This behavior was associated with good mode coverage of diffusion mapping, which requires a high Lipschitz constant. However, the previous experiments in the literature were limited to just one architecture, namely Stable Diffusion 1.5. The presented work examines a wider range of models and architectures and discusses the effect of distillation and few-step generation in Wan2.1, FLUX Klein distill, and Z-Image Turbo. Additionally, we present an analysis of the internal state of DiT during denoising to determine specific layers contributing to the high Lipschitz constant of a total mapping.