Nano Banana 2.1's consistency claim faces a 30-edit test that shows drift

Google Flow says its Nano Banana 2.1 image model keeps a subject consistent after repeated edits, citing a single social-media test by @pleometric that reconstructed one image 100 times with the prompt "reconstruct this exactly as it is" and found 2.1 preserved the original subject best.

That test rests on one user, one prompt, and one setup, and has not been independently verified.

A later test by Artificial Analysis gives a more mixed picture. Running four editing models through 30 consecutive edits of one real estate photo, each editing its own previous output, it found Nano Banana 2.1 made its edits locally but shifted and gradually darkened the rest of the image. Ideogram 4.5 and FLUX 3 left 95% or more of the image essentially untouched on small edits, and GPT Image 2.5 Sunburst re-rendered most of the image each time.