You can achieve these higher resolutions, fidelity and close to 100% reproduction of objects with considerable effort in generative AI right now, it's even a very enjoyable experience to do so, but not with a closed model like Nano Banana. When Google or other major image model providers give us proper fine tuning and custom inference, particularly for applications like upscaling, then we can have a serious conversation.
At present, while I understand the fascination, everyone should take a step back and consider what we're actually trying to replace. We are stuck at these low inference resolutions since SDXL. For genuine high-end commercial applications, we're not there yet, and we won't be without another significant leap in GPU technology.