Nano Banana 2.1: posters, lettering and grounded images from text
Nano Banana 2.1 is a strong text-to-image model for design work and clean lettering, with optional web search. How to prompt it and where it slips.
Ask an image model for a poster with a headline and you usually get a nice picture with something that looks like letters. Look at the travel poster among the examples of Nano Banana 2.1: the headline is spelled right, set in a clean typeface and placed where the prompt asked for it, and so is the festival line at the bottom. That is the short version of why this model is worth knowing: it treats layout and lettering as part of the job, not an afterthought.
What it is good at
- Design work. Posters, covers, menu boards, flat illustrations with a clear grid. It was tuned for visual design, and it shows in how it balances shapes and empty space.
- Following a long prompt. Prompt adherence is one of its stated strengths: when you list five objects and a palette, you tend to get five objects and that palette.
- Keeping a subject consistent across a series, when you describe it the same way each time.
- Photo looks too. The examples on its page include a candid portrait, a rainy street at dusk and an overhead food shot.
The settings that matter
Besides the prompt there are only a few. Aspect ratio covers the usual shapes plus some extreme ones, all the way to 1:8 and 8:1, which is handy for banners and tall story strips. Resolution has three steps; the smallest is the default and is right for testing. Output format is PNG by default, with JPEG as the lighter option.
The two advanced toggles are the unusual part. Web search and image search let the model look things up while it draws, so a prompt about something recent or specific has real information behind it. Both are off by default.
Three prompts that play to its strengths
A mid-century travel poster for a fictional mountain town called ALTAVERDE. Bold flat shapes: green slopes, a red cable car, a pale yellow sun. Headline at the top reading "ALTAVERDE", a smaller line at the bottom reading "Winter Lights Week". Clean grid, subtle paper grain, limited warm palette.A chalkboard cafe menu on a brick wall, three sections titled "Coffee", "Tea" and "Cakes", each with three hand-lettered items, small chalk drawings of a cup and a cake in the corners, warm morning light.Overhead photo of a rustic breakfast table: sourdough toast with smashed avocado and poached eggs, a cup of black coffee, fresh berries, linen napkin, soft morning window light, food photography.Lifehacks
- Test small, finish big. Draft at the default resolution until the layout is right, then run the winner once at a higher step. The Generate button updates its price when you change the setting, so you always see what the big run costs.
- Write the text you want and the text you do not. The model likes to fill a poster: in the example poster it added a tagline under the headline that the prompt never asked for. If you want only your words, end the prompt with "no other text".
- Switch on web search only when it helps. For a fictional town it adds nothing. For a real landmark or a current product shape it can make the result more accurate.
- Pick PNG when you will edit. If the picture goes into a design tool for more lettering, PNG keeps the edges crisp.
Who should pick something else
If you are still exploring ideas and do not need lettering, a draft model like Fast Image is a much cheaper sketchbook. If you are editing an existing picture, the image to image models on the models page are the right tool. And if you want a lot of empty space for your own text instead of the model's, read how to leave room for text first.
The honest limitation
Clean lettering is not perfect lettering. Short headlines come out well; long paragraphs, tiny print and unusual words still get letters swapped or invented, as with every image model today. Check every word before anything goes to print, and for body copy, add the text yourself in a design tool. The hands and text guide has the fixes.
Questions
Can Nano Banana 2.1 write readable text in images?
Short headlines and labels usually come out clean. Long passages and small print still have mistakes, so proofread and add body copy yourself.
What does the web search option do?
It lets the model look up real-time information while generating, which helps with real places and current things. It is off by default.
Which resolution should I choose?
Draft at the default, the smallest of the three steps, and render the final at a higher step once the layout is right.
