>>3021842SDXL was trained at 1024x1024 resolution so expects anything in the prompt to fit in about that space. If you go too far beyond that then it will start to duplicate things or cause other issues because the AI thinks there should be more to fill the space.
Like
>>3021870 said, typically the way to get higher resolution images that look good is to gen at something around 1024x1024 and run an upscaling process after the initial gen, like the Hi-res fix feature built into A1111/Forge/Reforge.
