The cross-tile challenges were quite robust - every model struggled with them, and we tried with several iterations of the prompt. I'm sure you could improve with specialized systems, but the models out-of-the-box definitely struggle with segmentation