Correct. Very well known artifact, you see it in all the GANs and diffusion models depending on the choice to do center vs random crop. Often changes results by several FID points (random crop adds a lot of variations and is much harder to model, confusing the NNs), which is why I suggested the cropping conditioning idea that SDXL uses with excellent results. (This is another example of how many of OP's observations are essentially minor, contingent aspects of the data or arch, and don't generalize beyond SD1.)