A 2 A.M. Codename and a Face That Finally Looked Like You

Stephanie Zhan and Pat Grady of Sequoia sit down on Training Data with Nicole Brichtova and Hansa Srinivasan, the product and engineering leads behind Google's Nano Banana image model. The hosts frame the punchline before the guests even speak: a 2 a.m. code name that became a cultural phenomenon, built on high-quality data, long multimodal context, and disciplined human evals. The comparison that makes the leap concrete is the old recipe. Two years ago you fine-tuned on ten pictures of yourself, waited twenty minutes, and still might not look like you. That cost and friction kept the capability out of the mainstream. Now the pitch is one photo, a sentence, a few seconds, and a conversation you can keep going.

The first community move that surprised Hansa was obvious only in hindsight. People immediately mixed Nano Banana with video models from several vendors, locking a face and a scene in stills, then feeding those frames into video so characters and settings held across cuts. The workflow is not fluid — users bounce between products, though some teams are trying to stitch the models together — but videos from before and after the launch look markedly smoother, with scene changes that feel closer to what a creator actually wants. She had not fully predicted that people would do this on day one.

Nicole's unexpected favorite is learning. She met someone generating sketch notes on technical topics, even though text rendering is not where the team wants it. He had hacked around the weakness with enormous prompts, then fed his father's university chemistry lectures into Gemini with Nano Banana. The notes came back coherent and visually digestible. For the first time in decades, father and son could talk about the father's work. People, she says, keep finding inputs the builders did not anticipate, and pulling performance out of the model that still surprises the people who trained it.

The aha moment was Nicole's, and it was a vanity prompt. In an internal demo she uploaded a photo of herself and asked to be put on the red carpet. It came back looking like her. She compared it with every prior model the team had; none of the others did. Colleagues nodded at the red-carpet shot and then took a couple of weeks to feel the same magic, because each person had to try it with their own pictures. That has become the main consumer gesture: seeing yourself in a new way, including the 3D-figurine meme of a computer, a toy box, and you as the figure, three times in one frame.

What was miles better was simple: it looked like her. Judging character consistency on a stranger's face is a weak test. Nicole can look at an AI version of a host and find it acceptable; the host will immediately see the parts that are wrong. The only reliable judge is the person in the photo, or someone who sees that face every day. The team now runs evals on many members' own faces for that reason. Hansa notices a large gap between scoring Nicole, whom she sees constantly, and scoring a random person in a set. Identity preservation is fundamental to the model being useful, and it is surprisingly tricky, which is why so many other systems still miss it.

It was an explicit goal, not an accident of scale. Past Google image models had a gap on consistency: when you edit, you want some parts to stay put and others to change, and earlier systems were bad at that. Advertisers had said for years that a product in a lifestyle shot has to look 100 percent like the product or it cannot run. The team thought it had the right architecture and data recipe; what surprised them was how good the finished model actually was. The intuition from phone editors and Photoshop makes the technical bar even steeper. A lay user expects the basic rule of editing — do not mess with what you were not asked to touch — and generative models find that shockingly difficult.

Pat wants a quantitative eval for "that's actually me." Hansa does not pretend face consistency is easy to score, for the reason Nicole just gave. Image generation in general has taught them that human evals are foundational. A dedicated team builds tooling and practices so people can judge subtle things — faces, aesthetic quality — that resist a number. The rest of the stack is unglamorous: eyeballing by different people, then community testing that starts inside Google and DeepMind with artists and executives. Quantitative benchmarks can say the new model is 10 percent better. They cannot carry the emotional story of seeing yourself in a new way, or of restoring a family photo you cut up at age five. Math and logic have a ground-truth answer and can be auto-evaled. Visual media is subjective. You need the qualitative feedback to tell that story.

Asked for the technical breakthroughs behind single-image consistency, Hansa starts with data that teaches the model to generalize, and with the fact that this is a Gemini model. It is a multimodal foundation system that has seen a lot of data. That generalization is the secret sauce: you need a model that already understands the world before you can spend that understanding on keeping a particular face still. Nicole adds the long context window. You can give one photo of yourself, or several, and on the output side you can iterate across turns and actually talk to the model. Two years ago the ten-image, twenty-minute fine-tune never took off because people do not have that many pictures of themselves sitting around, and it was too much work.

As Gemini gets better, the image model inherits the multimodal window and the ability to hold context over a long conversation. The rest is attention. A lot of the gains come down to someone on the team being obsessed with a narrow problem; they have people obsessed with text rendering, so text rendering keeps improving. Hansa pushes back on the idea that quantity of data is the whole story. There are small design decisions at every point, and the quality of the data and the selections matters. Nicole names that unfashionable layer: the craft of AI, which the industry does not talk about enough and which she thinks is super important.

The core modeling team is small. Shipping took a village. Because the model went out across many products, close collaborators across surfaces easily get you into the dozens and hundreds, plus infrastructure teams optimizing every layer of the stack to survive the demand. Internally they joked that it took a small country. Design-wise they mixed personas with capability. Before training they know which abilities they want, and inference speed is itself a persona choice. This model is a conversational editor, so it had to be snappy: you cannot really talk to a system that takes a minute or two per image. That is the advantage stills have over video. From day one it felt consumer-centric, a way to put pro-level capability behind a text prompt, even though developers and enterprises ended up using it too.

The philosophical shift from the old Imagen line is that generalization is now the foundation, not a side effect of a pure image generator. Some abilities are specified — represent people, edit a photo, keep it looking like them. Others emerge from a model that reasons about visual information. People drop in a drawing of a geometry problem and ask for the angle; the model renders the solution. Nobody trained a geometry-tutor feature. It showed up because reasoning, math, and visual understanding landed in the same system. The family around Gemini still includes specialists: Imagen for images, Veo for video. The north star is one model that takes any modality in and emits any modality out, and they are not there yet. Image stays a little ahead of the curve because one frame is cheaper to train and serve; developments on stills, Nicole expects, show up in video about six to twelve months later. Veo 3 put audio into video generation; Genie 3 let you navigate a world in real time. Those specialized models are a testing ground. Over time Gemini should do all of it.

Pat has to ask about the name. It is easy to remember and distinct, and he suspects it gave the launch a boost. Nicole and Hansa call it a happy accident. The model went onto LMArena, where unreleased systems get code names and users compare two anonymous replies. They were going live around 2 a.m. Someone pinged another PM on the team, Nina, for a name. She was exhausted. Nano Banana was the stroke of genius that arrived in that state. Nicole is careful to say it was not her; it was a teammate. It was fun, easy to pronounce, and it came with an emoji, which she calls critical for branding. Nobody overthought it. Once it was live, everyone just went with it. It felt Googley and organic, then looked like marketing genius after the fact.

They leaned in. Bananas are now all over the Gemini app, because people were publicly complaining they could not find the model. Hansa has had Googlers ask how to use Nano Banana; her answer is that it is Gemini, just ask for an image. Google has been a consumer company from the start, and the silly name played to that reputation. Underneath the joke is a thesis Nicole states outright: fun is a gateway to utility. People come in to put themselves on the red carpet or live out a childhood profession. Once they are in the app, they start studying, solving math problems, learning something else. Having a little fun, in the name and in the product, is undervalued because it gets people in and then shows them what the models are actually good at.

Hansa's parents and their friends started using it because it felt easy and unintimidating. Chatbots have already broken some of the fear around AI, mostly for younger people. A fun, unserious entry point brought her parents' generation in. Her mom made silly images, had a great time, then realized she could remove people from the background of photos. Silly turned practical, then people noticed the model could give them diagrams and help them understand things. There is a real accessibility component in not making the first step feel like technology.

On the product side, Nicole sees three jobs. For consumers, a lot of Nano Banana prompts are a hundred words long, and people copy-paste them into Gemini because the payoff is worth the work. That prompt-engineering phase has to end; a consumer product should not require users to write like power users. For professionals, the need is precise control, robustness, and reproducibility. The model is already good at edit consistency and at not changing pixels, but it is not at 100 percent, and professionals need 100 percent — possibly even gesture-level control over every pixel. The third job, the one that excites her most, is visualizing information. The chemistry sketch-notes hack should work for anything. People are visual learners, and LLMs have not exhausted their potential to digest material as a diagram, an image, or a short video for a biology class. Right now about 95 percent of what these models emit is text. That is useful, and it is not how people consume information in the world.

Chatbots are an easy entry point because you do not have to learn a new UI. For visual modalities they start to feel limiting. Nicole wants a new visual creation canvas that does not overwhelm people as the models can do more and more, and as open-ended tools make it hard to explain constraints. Google Labs, led by Josh Woodward, is the internal group that takes frontier models and asks what entertainment, creation, and productivity should look like; they have shipped NotebookLM and, on video, Flow. She is excited that Flow might become a place where some of this creation lives. In the short term Hansa wants the model to work the way you expect every time, not just a lot of the time, and to sand down the small inconsistencies. In the long term, rich multimodal generation should mean Gemini does not default to a text explanation unless you ask for images. The learning platforms that took off over the last ten or twenty years — Khan Academy starting on YouTube, Wikipedia full of pictures, diagrams for every math lookup — should become the natural flow, which again requires seamless generalization between modalities.

There is also a spectrum of how involved the user wants to be. Nicole started as a consultant and still makes slide decks. Updating stakeholders is a case where you do not want to be in the weeds: you dump meeting notes, a few bullets, an old deck, and you want Gemini to go off, format it, and make visuals — agentic behavior. Designing your own house is the opposite. You want to be in the weeds, playing with textures and colors, asking what happens if you remove a wall. A good product has to serve both ends: fine-grained control when the user is creating, and a model that understands the request, anticipates the outcome, and does the intervening work when the user wants to walk away. It is close to hiring a designer today. You give a spec; they do the expert work. The models should be able to do that, and in many domains they still cannot.

Pat asks where the next competitive battleground is. Nicole splits it in two. Capability still matters: a single model that takes anything and turns it into anything else is a thing nobody has really figured out. Adoption, though, will be decided as much by interfaces. The industry still leans on chatbots, which are a great entry point and not useful for everything. The work is to get specific about who the user is, what they are trying to do, and what product you wrap around the model. Asked whether the frontier will still be moving this fast in five or ten years, she laughs. Five to ten years feels like twenty. Two years ago she would have said the space was moving quickly. Today she says it is moving faster than it was two years ago.

Capability also raises the deepfake question. Nicole describes an evolving balance: give people creative freedom and control without feeling overly restrictive, and still prevent the worst harm. Every output carries a visible watermark that it was generated with Gemini. Image, video, and audio also embed SynthID, an invisible watermark, which is the Google standard across Imagen and Veo on every product surface. As models get more capable, internal testing and external partners keep finding new attack vectors, and the mitigations have to keep up. Hansa calls the balance very hard. People will use the tool in good faith and in bad faith. Users are responsible for what they do, but SynthID is the technology that lets the team release these capabilities with some faith they can still verify content and fight misinformation.

Barred from a five-to-ten-year horizon, Pat asks for one to three years: what will be possible that we only dream about, and how will life change. Nicole wants personalized tutors and textbooks. There is no reason two people with different learning styles and starting points should share one book, yet that is how school is set up. An LLM tutor that notices you like basketball and explains physics with basketball analogies feels achievable, if the factuality bar is high and the system is grounded in real-world content. Asked whether that could be a Google product, she says somebody should look into it. Hansa's answer is personal throughput. She is getting married and made save-the-dates with the model; code assistants and models that filter and analyze data have already changed how the builders work. What she can finish this year versus two years ago is an order of magnitude more, and that is true of tech, not of most other industries, because the workflow integration has not happened. Friends in consulting still spend two hours moving logos around. In a year or two she wants that time back on the content and the client.

Startups, they argue, should live on the application layer Google is not trying to own. Creators already jump across four tools: ideate in an LLM, iterate key frames in an image model, generate in a video model, then do sound, music, and heavy editing in traditional software. Workflow products will spin up for that stack and for other verticals — consultant slide and pitch decks, sales, finance. Big companies may not go into those niches; startups are in a position to learn a specific client and automate the tasks. Hansa's closing note is how many aunts, uncles, parents, and friends the model has moved. They already used chatbots for information, including her mother's health questions, but visual media excites people in a way that is fun, intuitive, and emotional, because so much of how humans experience life is visual. Stephanie's three-year-old tied a fraying dog leash over himself like a warrior cape; she photographed him into a superhero. Her husband uses Google's Storybook to turn playground incidents into stories whose characters look like the boy, the parents, the dog, and his sister. Nicole's line for that future is the one that opened the episode: when cameras became accessible, anyone could capture reality. These tools capture imagination — they get what is in someone's brain onto paper visually, including stories that would only ever have been made for one or five people.