By skyforbes I noticed with the new Nano Banana 2 that it's great at generating images with text BUT it seems to really struggle with variations of those images. For example, you can see below that it was able to generate a Coca Cola can with the word "Spine" on it, but when trying to turn the can yellow or make the logo more realistic, it took the word spine literal.
Testing on Nano Banana 1 by branching off the original image generated by Nano Banana 2, it had no issue following the prompts and maintaining the text.
This might be the only solution I'm aware of by using this template (or building your own) link the two models to get your desired output.