Nano Banana 2.1 draws beryllium orbitals, but the GPT Image 2.5 comparison is a mismatch
The author says Nano Banana 2.1 can draw beryllium's quantum orbitals fairly accurately after two revisions, and argues that comparing it with GPT Image 2.5 is a mismatch since realism remains the strength of Pro and GPT Image 2.5.
Nano Banana 2.1 can (quite) accurately represent the quantum orbital structure of the Beryllium atom. Can't one-shot it but with the right scientific guidance it can move the needle in the right direction (attached image was after 2 revisions by me). Also, I keep seeing @NanoBanana 2.1 lined up against @OpenAI's GPT Image 2.5, and that's the wrong fight for a cheap Flash model. And to be honest, @Google invited some of that by saying 2.1 "beats all of its previous models". I don't think it does in real world use, even if it seemingly does on @arena's text-to-image leaderboard, at least for now. It has a preliminary score of 1328, ahead of Nano Banana 2 at 1261 and Nano Banana Pro at 1248, and it costs about $0.034 per 1K image, roughly half what Nano Banana 2 charges. That's a solid gain for the price. One should understand that the realism is the job for Pro. On the photorealistic board 2.1 sits at #4, about 90 points behind GPT Image 2.5, again on preliminary numbers, and Pro hasn't been updated since last November. I hope the next Nano Banana Pro comes out considerably better than what shipped today.