Title: Entropic Markers of Class Speciation within Diffusion Models

Abstract:

Diffusion models do not uniformly reconstruct semantic structures throughout the generation timeline. Rather, generated samples shift from a state of semantic ambiguity to distinct class commitment during a specific, narrow time window. While recent theoretical studies attribute this shift to dynamical instabilities occurring along directions that separate classes, there remain few practical techniques for identifying or leveraging these critical intervals in pre-trained models. This study demonstrates that monitoring the class-conditional entropy of a latent semantic variable—given the current noisy state—offers a dependable indicator of these transition phases. Furthermore, by confining the entropy calculation to semantic partitions, the metric can distinguish semantic decisions across varying degrees of abstraction. We examine this phenomenon using high-dimensional Gaussian mixture models, revealing that the entropy rate aligns with the same logarithmic time scale as the speciation symmetry-breaking instability previously documented in variance-preserving diffusion. Our approach is validated on EDM2-XS and Stable Diffusion 1.5, where class-conditional entropy reliably isolates the noise regimes essential for the development of semantic structure. Additionally, we employ this framework to measure how guidance mechanisms redistribute semantic information over time. Collectively, these findings bridge information-theoretic and statistical physics viewpoints on diffusion processes, establishing a rigorous foundation for time-specific control strategies.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC