Google launches Nano Banana 2 with new features, performance benchmark, and usage

Google’s Gemini 3.1 Flash Image, codenamed Nano Banana 2, expands on the “Pro‑grade” visual reasoning of its predecessor while slashing latency to a “Flash” experience, according to the CometAPI2025 report posted on March 1, 2026. The model supports full‑resolution 4K output, aspect‑ratio control, and multi‑modal generation (image + associated text), positioning it for both lightweight assets and production‑ready visuals. Its core capabilities include single‑shot and multi‑step text‑to‑image generation, in‑painting, multi‑image fusion, and iterative prompt‑steering that let developers “sketch → refine → finalize” within a single pipeline. A notable addition is character‑consistency tracking, which preserves facial features and stylistic cues across successive edits—a feature the report highlights as essential for storyboarding and serialized art production. Finally, Google embeds SynthID watermarks in every output, providing provenance metadata that aligns with the company’s broader transparency agenda.

In blind side‑by‑side testing on the GenAI‑Bench Human Elo Evaluation, Nano Banana 2 outperformed every competitor. The model posted an overall preference Elo score of 1079 ± 7, a six‑point lead over the earlier Gemini 2.5 Flash Image (1073 ± 5) and a 58‑point margin against GPT‑Image 1.5 (1021 ± 5), the report notes. Visual‑quality Elo scores tell a similar story: Nano Banana 2 achieved 1140 ± 6, eclipsing the 2.5 Flash variant by 11 points and beating GPT‑Image 1.5 by 97 points. The data suggest that the incremental architectural refinements—presumably larger diffusion steps and a tighter text‑image alignment module—translate into measurable gains in both user preference and perceived realism.

The model’s performance gains are matched by a suite of production‑grade controls that address the cost barriers that have kept many enterprises away from generative image pipelines. VentureBeat’s Sam Witteveen points out that Nano Banana 2’s “low‑latency generation” and “high‑fidelity image understanding” together reduce the need for costly post‑processing and manual touch‑ups. By allowing developers to supply reference images and issue natural‑language edit commands, the API eliminates the iterative back‑and‑forth that typically inflates compute budgets. The report also emphasizes that the model’s ability to retain subject consistency across edits cuts down on re‑render cycles, further driving down operational expenses for large‑scale content creation.

Google has baked Nano Banana 2 into the broader Gemini ecosystem, making it the default image engine for Gemini’s APIs, Airtable integrations, and Google Cloud services. The Verge confirms that the model is now available to free users, extending its reach beyond paid tiers and giving developers early access to the new feature set. Wired’s hands‑on coverage notes that the model’s 4K output and SynthID watermarking are already being leveraged in real‑world workflows, from marketing asset generation to rapid prototyping of UI mockups. By unifying high‑quality generation, editing, and provenance in a single service, Google aims to lower the friction point that has historically separated research‑grade generators from enterprise deployment.

From a technical perspective, Nano Banana 2 represents a convergence of two previously divergent design goals: ultra‑low latency and professional‑grade visual fidelity. The CometAPI2025 analysis attributes this balance to a “strategic evolution” of the underlying diffusion architecture, which now incorporates a higher‑capacity transformer encoder for text conditioning and a refined decoder that can produce 4K frames in sub‑second intervals. The model’s iterative workflow support—allowing interleaved text prompts and image inputs—suggests a shift toward “prompt‑programming” paradigms where developers can script complex compositional tasks without external orchestration. If the benchmark results hold across broader user bases, Nano Banana 2 could set a new baseline for generative image services, compelling competitors to match both speed and quality in a single offering.