Google Drops Nano Banana 2: The New King of AI Picture Technology? – Decrypt

Google has been releasing AI software program at a staggering tempo recently. Within the final week or so alone, we have seen Gemini 3.1, Lyria, and Pali, which got here with a photograph shoot function that turned out to be a real crowd pleaser. And now, the follow-up to arguably the largest picture era hit of final 12 months has arrived.

Nano Banana 2, launched Thursday, “brings the high-speed intelligence of Gemini Flash to visible era, making speedy edits and iteration potential,” Google mentioned in an official weblog submit, including that “it makes once-exclusive Professional options accessible to a wider viewers.”

Here is the fast breakdown. The unique Nano Banana was really named Gemini 2.5 Flash Picture, and was principally that: A picture generator primarily based on Gemini 2.5 Flash. Then Nano Banana Professional got here alongside, which was Gemini 3 Professional Picture, and it turned the gold commonplace for AI picture enhancing when it launched final November.

Nano Banana 2 is technically Gemini 3.1 Flash Picture—so it is not a direct sequel to Professional, however extra like a considerably upgraded model of the unique, now operating on the newer Gemini 3 Flash spine. Complicated? Sure.

The pitch right here is straightforward: take every little thing that made Nano Banana Professional particular, and make it run at Flash pace.

The brand new Nano Banana 2 rolling out at present throughout Google’s ecosystem. Within the Gemini app, it replaces Nano Banana Professional because the default throughout Quick, Pondering, and Professional fashions. Google AI Professional and Extremely subscribers can nonetheless entry Nano Banana Professional for specialised duties by regenerating through the three-dot menu.

It is also reside in Google Search’s AI Mode and Lens, out there through the Gemini API in AI Studio and on Vertex AI in preview, and it is the brand new default picture era mannequin in Circulation at zero credit for all customers. Google can be increasing SynthID watermarking and including C2PA Content material Credentials assist to present platforms higher instruments for figuring out AI-generated media. The SynthID verification function has already been used over 20 million instances since November.

What’s new in Nano Banana 2

The most important headline is world data. Nano Banana 2 can pull from real-time net search throughout picture era, which implies it will probably render particular topics with accuracy. Logos, landmarks, latest occasions, model identities—it is aware of what issues appear like as a result of it will probably look them up, not simply guess from coaching knowledge.

Textual content rendering acquired a severe improve too. Now you can generate correct, legible textual content inside pictures, whether or not you are spelling it out within the immediate or letting the mannequin determine what to put in writing primarily based on context. It additionally handles in-image translation, so you possibly can localize an advert marketing campaign throughout a number of languages with out rebuilding the visible from scratch.

Topic consistency is pushing into new territory too. The mannequin can keep character resemblance throughout as much as 5 topics, and preserve the visible constancy of as much as 14 objects in a single workflow in response to Google. That is a giant deal for anybody constructing narratives, storyboards, or constant model belongings.

On the manufacturing facet, you get every little thing from 512px all the best way as much as 4K, with native assist for a variety of side ratios. Instruction following can be tighter than in earlier Flash fashions, which in observe means fewer prompts that type of do what you requested, and extra prompts that truly do precisely what you requested.

The reasoning can be now configurable. Builders can set considering ranges from Minimal (the default) all the best way to Excessive or Dynamic, letting the mannequin motive by way of advanced prompts earlier than committing to a render. That mixture of pace and non-compulsory deliberation is the place numerous the standard features are coming from.

Testing the mannequin

The pace claims are actual. We requested Nano Banana 2 to generate an entire Bitcoin ecosystem timeline, together with analysis and closing art work. The total course of took roughly the identical period of time Nano Banana Professional wanted simply to finish the Bitcoin timeline alone. Once we adopted that up with an Ethereum timeline immediate, it barely registered as extra time. That could be a significant hole for anybody operating iterative pipelines or constructing at scale.

The world data functionality genuinely adjustments how the output feels. Once we prompted for a historic crypto timeline, the mannequin searched a number of sources, chosen probably the most related occasions, and structured the artwork round them. It wasn’t generic. The mannequin made editorial selections. The one actual flaw we noticed was a lacking visible hyperlink between the top of 1 part and the beginning of one other. Every thing else holds collectively. Nano Banana Professional, by comparability, produced one thing extra generically inventive and made no obvious effort to supply or prioritize occasions.

For instance, that is what Nano Banana 2 generated when prompted “Create a timeline of Bitcoin’s historical past, highlighting crucial occasions from its creation to at present. widescreen, youngsters drawing fashion” utilizing considering.

For distinction this is identical era utilizing Nano Banana Professional:

Character consistency and textual content dealing with had been probably the most spectacular components of our check outcomes. We requested the mannequin to generate {a magazine} entrance cowl, and each line of textual content got here out correct and properly outlined. No garbled characters, no drifting typography.

Nano Banana Professional can be robust right here, but it surely produces extra hiccups, and its journal cowl output had a 3D render high quality to it that comes throughout as artificial.

Nano Banana 2’s outcome seems to be photorealistic. It additionally reveals fewer garbled characters total when producing textual content by its personal reasoning, not simply when explicitly informed what to put in writing.

That mentioned, the mannequin has a transparent content material ceiling. We requested Nano Banana 2 to edit an actual picture and alter the outfit to underwear. After a protracted reasoning cycle, it refused. That’s to be anticipated, had been it not for the truth that it refused to generate the edit on the picture of a girl, however not on the picture of a person.

Asking for a swimsuit swap labored nice. The censorship stage seems roughly equal to Nano Banana Professional, which implies something pushing towards specific territory or manipulation of actual individuals in suggestive contexts will get blocked. This issues greater than it’d sound, and we’ll get to why in a second.

Seedream 5: Nano Banana 2 has competitors

Here is the factor about launching a flagship picture mannequin in late February 2026: ByteDance launched Seedream 5 the exact same week.

Seedream has turn out to be a group favourite during the last 12 months, and for good causes. It is versatile, it is cost-efficient—round $0.035 per picture through the API which is round a 3rd of Google’s costs—and its content material moderation is significantly extra permissive than Google’s. That final level has constructed it a loyal following amongst creators who want extra room to work with actual individuals or push visible boundaries.

Seedream 5 brings real-time net search into its era pipeline, improved reasoning, stronger reference consistency, and assist for as much as 14 reference pictures in a single multi-round enhancing workflow. It generates at 2K and 4K in seconds. It might probably additionally run domestically, which Google doesn’t permit, and is obtainable in ByteDance’s CapCut and Jianying, and thru the usual API.

Briefly, each Google and ByteDance launched web-search-grounded, reasoning-enhanced picture fashions in the identical week. That tells you one thing about the place the entire class is heading.