Global selection, local completion: a probabilistic anatomy of diffusion U-Net

Two studies examine how diffusion U-Nets model complex images efficiently. Coarse-to-fine conditioning makes fine detail predictable through small, local models, reducing effective dimensionality. At the bottleneck, sparse active channels encode global image information whose geometry reflects semantic similarity even without labels or text conditioning. Self-guided reconstruction probes that representation, supporting an account that combines compact global selection with locally conditioned completion.