Google absolutely COOKED! nano_banana is Gemini, & they just won image gen.

Google’s long-hyped “nano_banana” image model has been revealed as Gemini 2.5 Flash Image Preview—a fast, editing-capable system that delivers unusually strong character consistency and prompt-following, while also posting benchmark wins against major rivals. The practical impact is straightforward: users can generate and edit images with “Photoshop-level” control inside Google’s AI Studio (with limited free quota), and then scale up via the Gemini API. For many testers, the combination of speed, consistency, and edit fidelity is the real story—especially when the edits preserve identity and scene details rather than replacing them with generic artifacts.

The model’s performance is framed through comparisons against GPT-4o’s native image generation (high-quality mode), Flux 1 (context max when image editing), and the older Gemini 2.0 Flash Image. Gemini 2.5 Flash Image Preview comes out on top in most categories, including overall preference, character handling, creative tasks, infographics, and object/environment manipulation. Stylization is the main area where competitors—particularly GPT-4o and “Quen image edit” in the transcript—hold an edge, but Gemini 2.5 Flash remains close enough that it still “crushes” Flux and the earlier Gemini 2.0 Flash Image in multiple comparisons.

Beyond benchmarks, the transcript highlights what users can actually do with the system. In one test, a prompt themed around banana-inspired armor produces an image in about ten seconds with consistent facial identity, a stable background, and a coherent suit design. Another example modernizes a vintage “uranium burger” photo: the model colorizes the black-and-white image and updates details like clothing, signage, and background elements while keeping the overall scene grounded. A more demanding edit places a car onto the moon with Earth in the background, with lighting and reflections adjusted to match the new environment—down to wheel and door-handle details—completed in roughly 35 seconds.

The editing strength extends to adding labels and glows around objects in a dog/pet-carrier photo, and to cinematic scene generation where the same person can be reused across consistent “movie-like” frames. The transcript also credits Gemini 2.5 Flash Image Preview with native image generation alongside editing: prompts ranging from a cathedral made of pulsing jellyfish to armored lemon mechs and surreal “dream-home” landscapes are said to land accurately, though the model hits detail limits when pushed toward extreme clarity or dense scenes.

Character consistency becomes a centerpiece in a “Story Book” experiment, where a hyperreal narrative about an abduction and the “singularity” is paired with consistent visuals of the same protagonist across multiple scenes. The transcript claims the storybook can be generated in about ten minutes, with the model producing coherent, sequential imagery that matches the written prompts.

Cost and availability are positioned as additional advantages: Gemini API usage is described as far cheaper than OpenAI’s native image generation pricing (about 4 cents per generation versus about 19 cents), and the service is said to be available in Europe from the start. The overall takeaway is that “nano_banana” is not just a flashy generator—it’s a fast, editing-first Gemini model that competes strongly on consistency and real-world usability, with Google’s broader Gemini ecosystem (including Notebook LM and other Gemini releases) presented as part of the same momentum.