Image editing program -> different versions of the image, each with some but not all of the elements you want, on each layer -> mask out the parts you don't need/apply mask, fill with black, soft brush with white the parts you want back in. Copy flattened/merged, drop it back into the image model, keep asking for the changes. As long as each generation adds in an element you want, you can build a collage of your final image.

It's the first thing I tried, because Nano Banana 2 deteriorates the output with each turn, becoming unusable with just a few edits.

ChatGPT Images 2.0 made it unusable at the first turn. At least in the ChatGPT app editing a reference image absolutely destroyed the image quality. It perfectly extracted an illustration from the background, but in the process basically turned it from a crisp digital illustration into a blurry, low quality mess.

There was an Edit button in one of the images in the livestream