Exactly. Notably, in my experiments, diffusion models based on U-Nets (e.g. SD1.4, SD2) are worse at capturing "correlations at a distance" like this in comparison to newer, DiT-based methods (e.g. SD3, PixArt).