Semantic information production peaks earlier for less visually similar CIFAR-10 class pairs

Handke, Koulischer, Raya & Ambrogioni recover the Bayes-optimal class posterior at each diffusion noise level from real class-conditional DDPMs (PixelCNN++-backbone U-Nets) trained from scratch on CIFAR-10, using the model's own conditional and unconditional noise predictions rather than an auxiliary classifier [handke-etal-2025-measuring-semantic-information-production-generative-diffusion] The resulting class-conditional entropy rate -- bits of class information produced per unit diffusion time -- peaks in an intermediate noise interval for every class pair tested, preceded by data-mean convergence and vanishing near the final denoising step [handke-etal-2025-measuring-semantic-information-production-generative-diffusion] Class pairs sharing less visual structure (deer vs. car) show this semantic-information-production peak occurring earlier in the reverse process than pairs sharing more structure (deer vs. bird, deer vs. cat), showing the entropy-rate trajectory's shape is class-pair-specific rather than a single universal curve [handke-etal-2025-measuring-semantic-information-production-generative-diffusion]