These are limitations of a single text-to-image gen, which is the least interesting way to use those models. When guided by a previous low-res generation, it wo

These are limitations of a single text-to-image gen, which is the least interesting way to use those models. When guided by a previous low-res generation, it won't fall apart at arbitrary resolutions, that's how all diffusion upscalers work. Just don't expect being able to fit every detail in one pass, use multiple ones (that's how detailers work).