Nano Banana 2 has an ace up its sleeve

Rendering correct textual content has lengthy been a stumbling block for even essentially the most superior AI picture turbines, nevertheless it’s among the many strongest fits of Google’s just-updated Nano Banana 2 engine.

Accessible now within the Gemini app (you’ll additionally discover it in Google Search, AI Studio, and different Google merchandise), Nano Banana 2 boasts a spread of latest options, together with as much as 2K decision that may be upscaled as much as 4K, “enhanced” instruction following that helps the mannequin adhere higher to your prompts, and the flexibility to lean on Gemini’s “real-world” data, permitting it to attract real-time data by way of net search because it renders pictures.

Not unhealthy, however much more spectacular is Nano Banana 2’s textual content constancy. I’ve been asking Nano Banana 2 to create pictures with billboards, indicators, newspapers, and different objects with embedded textual content, and it’s been performing like a champ, largely avoiding the gibberish that earlier AI picture turbines sometimes produced when making an attempt to render letters and phrases.

For instance, I prompted Nano Banana 2 to render a picture of a robotic smoking a cigarette in Instances Sq., with a neon marquee studying “Nano Banana 2 on Broadway” within the background. No downside, and it rendered the picture (above) in roughly 10 seconds.

I then requested Nano Banana 2 to create a photograph of a lady studying a newspaper in a breakfast nook, with the newspaper headline studying “Nano Banana 2 makes its debut.” However for this take a look at, I upped the ante: I requested the engine to jot down the sub-headline and the article itself, and directed that the story ought to particularly be about Nano Banana 2. 

Nicely, the mannequin bought the subheadline good, however even higher, it did write the article–up to some extent, anyway. The article textual content is a tad wiggly, however you possibly can virtually learn it.

I then pushed Nano Banana 2 a little bit extra, asking it to zoom in on the article and improve the textual content.

Google

Right here, the textual content rendering broke down a bit, “Google has unveiled its newest akthrough [sic] in generative AI, the ‘Nano Banana 2’,” the article reads, “promising a serious leap [the word “leap” is partially obscured by a finger] in picture era constancy.” Not unhealthy, however as you retain studying, the textual content constancy does begins to crumble.

Lastly, I attempted asking Nano Banana 2 to attract a diagram of–effectively, itself. “Render a diagram of nano banana 2’s structure throughout the better Gemini framework, full with textual content captions,” I prompted, and about 15 seconds later I bought this:

Google

Wanting carefully on the diagram, I didn’t see any textual content gibberish in any respect, and the diagram and captions appeared to make sense, or at the very least it did to my untrained eye.

Plugging the diagram into the Gemini app, the “considering” model of Gemini assured me it was a “remarkably correct architectural map” of the general Gemini framework, precisely depicting how the brand new mannequin can deal with as much as 5 constant characters inside a picture workflow. It additionally appropriately referenced the brand-new GemPix 2 Diffusion Renderer, the Nano Banana 2 element that takes the engine’s native 2K picture renders and upscales them to 4K.

All in all, very spectacular, though Nano Banana 2 additionally begs the query of when OpenAI will counter with a follow-up to final 12 months’s GPT Picture 1.5. That might occur any day now, if not at present.