A Name Invented at 2:30 a.m. Put Gemini on Top of the App Store

ChatGPT had sat at the top of the App Store's free chart almost since the day it launched. In September 2025, Gemini finally knocked it off. What drove the surge, Jacob Effron says on Unsupervised Learning, was an image model with real gains in character consistency and quality. Google's official name is Gemini 2.5 Flash Image. The internet, and eventually Google's own CEO posting banana emojis, called it Nano Banana. Effron sits down with Nicole Brichtova, on the product side, and Oliver Wang, on research.

Before release, the use case that most excited Brichtova was character consistency: seeing yourself in a new scene. She filled slide decks with her own face on wanted posters, as an archaeologist, in childhood dream jobs. The team later built an eval set around her face and colleagues' faces and eyeballs every new checkpoint against it. Effron calls having an eval named after your face the ultimate honor in AI. After launch, people turned themselves into figurines. The surprise, though she says it should not have been, was colorizing old photographs: seeing yourself as a baby, or parents as young people, in images that had only been black and white. The model also turned objects in a photo into holograms, something they never trained for.

The loudest requests on Twitter are professional: resolution above the current 1K cap, transparency, and better text rendering. Wang is more interested in a different frontier. The model can now take a fuzzy ask. Redecorate my room, I do not know how. It suggests palettes that fit what is already there. The leap he wants is to pour language-model world knowledge into pictures, so an image can help you decide, or answer how something works. Aesthetics and personalization still sit far from that. He hopes personalization stays at the prompt and context layer, feeding the model photos of your closet rather than serving everyone their own weights, which "sounds like kind of a nightmare." Aesthetic control beyond prompting, if it is needed, will probably live in pro workflows.

What made the model so much better? Wang's answer will disappoint anyone hunting for a secret. It is a boring case: not one specific thing, but all the details, a tuned recipe, and a team that had been on the problem for a long time. Even they were taken aback by how far it went. They knew they had a cool model and were eager to ship. On LMArena the Elo scores were high, a sign the model was useful. The metric that told him it was special was people showing up at LMArena specifically to use it. The team had to keep raising queries per second. They had not expected that.

Image models, he says, benefit almost 100 percent from the world knowledge of language models. Gemini 2.5 Flash Image is a Gemini model. You talk to it as you talk to Gemini, and it understands what Gemini understands. That integration was a step-function in utility. Two or three years ago you had to specify every attribute. The old hack was to expand one sentence into a ten-sentence prompt. That is no longer required, and it is not magic prompt rewriting on the backend. The models themselves got sophisticated enough to get it.

The name itself is almost an accident. A PM on the team, Nana, was up at 2:30 in the morning on the release and coined it. It stuck because it was fun, then stuck as a semi-official name because Gemini 2.5 Flash Image is harder to say. Effron's branding takeaway: your name should have an emoji. Hugging Face figured that out first in this industry. Wang wonders aloud how much of the hit is simply that people like saying Nano Banana.

Users split in two. LMArena regulars and developers know what they want the second a model drops, and they invent uses the team did not expect. Gemini users often hit a blank box. Making it easy for them is the product job. When people heard about Nano Banana and opened the app, they could not find it, so the team put banana emojis everywhere. They partnered with creators on seeded examples that deep-link back with the prompt already filled. Still missing, Brichtova says, is visual guidance in the zero state, and gesture editing: erase an object the way you would on a scratch pad, instead of writing a paragraph. Long prompts are not a natural mode for most consumers.

Her bar is the parent test. If her parents can use it, it is probably good enough. They are not there. A lot of the remaining work is showing, not telling: examples people can reproduce, and sharing that is frictionless. Social spread helps the blank-slate problem because the model is personalizable by default. You see someone else's trick and try it on yourself, a friend, or a pet. Longer term she wants modalities to blend, with an interface that picks the right one: voice when talking is natural, a gesture when you need to point, an image when the answer should be visual. Nobody has really cracked how voice should show up. The hard parts are detecting intent, switching modes, and showing what is possible without dumping the user back onto a blank slate.

Chatbot users expect the bot to do anything, because it talks like a person. Explaining limits is hard when the tool can now do so much, and whatever you teach goes stale within three months as capabilities move. In the quickfire, Brichtova calls prompt-to-production-ready the overhyped idea. Social posts still hide a lot of work. Underhyped: the UIs of the future, how you show people what is possible and make it useful in a specific workflow. She has not seen a product that feels like the answer. She likes node-based interfaces and admits that is a specialized taste.

Evals for images are a harder bone than legal question-answering, Effron notes, because the answer is subjective and you can even argue about which direction to climb. One new loop helps. As language models and vision-language models get better, their intelligence can grade their own generations, a virtuous circle on both axes. Wang still puts people at the center. Users writing their own prompts on LMArena is the best evaluation, because they are the arbiters of the pictures they want.

Taste still matters. Brichtova says Wang will not admit it, but he is one of the people on the team with a really good eye for whether an image is working and where it fails. A few such people do the first pass on training outputs. Internally they call it eyeballing, "a very technical term." Feedback from X goes into the eval set both ways: working cases so they do not regress, broken cases the community wants fixed. She asks people to keep tagging them. If something sucks, Wang says, tag him. It is useful.

Wang spent most of his career in video, so his favorite advanced use of Nano Banana is pairing it with Veo 3. Iterate shots in image space the way film still starts with storyboards, then spend video-model compute once the frames are right. Architecture workflows surprised him: blueprint to something that looks 3D but is not, then to a design you can iterate, compressing the tedious stretch so people stay on the creative part. The same missing step shows up in code. Going from a prompt straight to a generated website always felt abrupt. Now you can iterate the UI in image space until the aesthetic is right, then generate the code. Garden redesigns on the team already work that way: preview in Gemini, then take the idea to a landscape architect.

That split is how they think about Gemini chat versus the API. Chat is the first step in ideation, almost never the final artifact. Pros and developers chain multiple models in visual-forward UIs. Chatbots get you started, give inspiration, and support fun, shareable uses with family. Users who want demanding workflows will keep leaving the chatbot. Editing from 95 percent to 100 percent depends on the person. Some users know what they want to the pixel and need Adobe-class tools, which already expose the API. Wang used to work there. Others want inspiration and looser constraints; a chatbot is enough.

An advertising detail Brichtova learned two days earlier makes the limit concrete. Which way a model looks in an ad changes the message. That control is very hard to get from a chatbot. Language is excellent for high-level ideas and a clumsy way to move something three pixels left. Both surfaces stay. Inside Google, Photos is an obvious home, because the library is already there: a family picture becomes a birthday card a few times a year. Education is another: explain photosynthesis to a five-year-old with a picture that does not exist on the internet. Google Slides is the one that makes both hosts, former consultants, light up. Photograph a whiteboard with columns and a headline, and stop spending a career on formatting.

Wang has been in the field since GANs, when a model could look decent on a front-facing face and fall apart if you turned the head. Diffusion models made people realize everything would change. Nobody predicted the speed. He credits friendly competition: Midjourney ahead for a long time because it nailed post-training for style and artistic look, narrowing generation to images that always looked good; Stable Diffusion, open-sourced, showing how large the builder community was; later Flux and GPT-class models widening the range without dropping quality, helped by better data and a natural scaling curve. The cost is that users got jaded in lockstep. People now complain about small flaws in pictures that, a year ago, did not even look real.

Effron asks whether image models have 10 percent of the road left, or whether in three years today's pictures will look funny. Wang is entirely in the second camp. Even ignoring new applications, image quality has a long way to go, and the name for that gap is expressibility. Some common things already generate perfectly, indistinguishable from reality. Leave the well-trodden prompts, ask for imagination or several concepts composed together, and quality drops fast. Best-case images today may look as good as best-case images in a few years. Worst-case images today will be significantly worse than the future floor. Broaden the model and people discover more uses. That is how it becomes actually useful.

Competition so far has let small teams hit state of the art, and he hopes that lasts. World knowledge, though, benefits from language-model scale, so the groups that can train large LMs may end up with the image models that know the most. Big Chinese labs are already a major player, as in language. Whether training on the best open-source LM is a disadvantage depends on the future of open source, which changes often. A year ago that looked like a safe path. Now it is less clear. Next year's pace, he hopes, is more: more smart people, more resources, acceleration on every axis. Image and video are very close friends. Techniques flow both ways, which is one reason video generation took off. Workflows stack: think in the language model, iterate frames because it is faster and cheaper, then go to video. Consistency is a single-frame problem in images and a multi-frame problem in video. Video's next list is the latest image-level control, resolution, temporal consistency, the same character across scenes, and longer coherent content. Omni-models that do everything may win in the long run. "Very close friends," he says, "and maybe we'll live together someday."

A frontier he thinks too few people watch is factuality. Infographics and a Niagara Falls photo with annotations look fine as demos until you zoom in: garbled text, repeated facts, errors. Language models walked the same path. GPT-1 and GPT-2 were toys for haiku. Nobody uses them that way now; they are for information and conversation. Images will move from creative tools to information tools, and someday people may talk to a video model when they need someone to talk to. Models should also get proactive. Search already chooses text, image, or both. If the query warrants a picture, the model should just produce one. Asked for a favorite Nano Banana artifact, both land on playing with their kids: funny locations, stuffed animals come to life. Click the bananas in the Gemini app or AI Studio, they say, and if something does not work, tag them on X.