Google's Gemini image model takes instructions the way a person would rather than the way a diffusion UI does: full sentences, spatial relations, and an explicit statement of what must not change. Three prompts that circulated publicly after the model was confirmed show three distinct patterns worth stealing. By the end you should be able to write in each shape and tell which one a given task needs.
The model behind the nickname
In August 2025, Google confirmed that the image model that had been circulating anonymously under the nickname "Nano Banana" was Gemini 2.5 Flash Image. It is reachable by asking Gemini to create or alter an image, through the Gemini API, and through third-party tools that wrap it.
The part that matters for prompting is architectural. This is a multimodal model that emits images, not a text encoder bolted onto a diffusion sampler. There is no negative prompt field, no CFG slider, no seed to fix. The prompt is the entire control surface, and it behaves like an instruction to an assistant that can see the picture: ordering matters, spatial prepositions work, and a conditional such as "if possible" is understood rather than blended in as tokens.
That is why prompt habits carried over from Stable Diffusion land badly here. Comma-separated tag soup gives this kind of model very little to act on. Sentences give it a lot.
Pattern 1: build a scene around an existing subject
turn this image into a character figure. Behind it place a box with the character's image printed on it and a computer showing the Blender modeling process on its screen. In front of the box, add a round plastic base with the character figure standing on it. Make the PVC material look clear, and set the scene indoors if possible.
Feed it a photo of a person or character and you get back a staged product shot of them as a collectible figure. What makes it work is not the subject, it is the layout instructions: behind it, in front of the box, on its screen. Every new element is anchored to something already established, so the model constructs depth by following a chain instead of guessing at a composition.
Two more details are worth copying. Material is specified separately from the object it belongs to ("make the PVC material look clear"), which stops the finish from being applied to the whole frame. And the least important requirement is phrased as a soft constraint ("if possible"). Ranking requirements inside the prompt gives the model somewhere to give ground when two of them conflict, instead of dropping whichever one it feels like.
The reusable skeleton:
turn this image into [object]. Behind it place [element]. In front of it, add [element]. Make [material] look [property]. Set the scene [context] if possible.
Pattern 2: change one thing, freeze everything else
Make the ramen bowl and ramen look like 2D whimsical hand drawn anime illustration keeping everything else in the image exactly the same
One sentence, and the closing clause carries most of it. Selective style transfer, where a single object is re-rendered in a different medium while the rest of the photograph stays photographic, is awkward to express with masks and trivial to express in words.
The freeze clause is the pattern. Without it, a style instruction is read as a global one and the whole image converts. "Keeping everything else in the image exactly the same" scopes the edit. Use that wording or something equally literal; vaguer versions such as "don't change much" hold far less well. Pair it with a precisely named target, note that this prompt names both the ramen bowl and the ramen, and you get a clean boundary without ever drawing one.
Pattern 3: cast and wardrobe as an explicit checklist
Portrait of a couple walking in a pedestrian street in Lisbon in the 1970s. She is wearing sunglasses, a white tank top and denim shorts with frayed edges. He is wearing sunglasses, a white shirt and jeans. She has long blonde hair. He has short dark hair.
This one is generation rather than editing, and its structure is a checklist: a scene sentence, then one sentence of wardrobe per person, then hair per person. Multi-subject prompts fail when attributes leak across subjects and she ends up in his shirt. Splitting attributes into separate sentences keyed to separate pronouns holds them apart far more reliably than one long comma chain does.
The period cue does double duty. "In the 1970s" sets clothing, film stock, colour palette and street furniture in a single short phrase. Naming an era is almost always cheaper than describing its visual consequences one item at a time, and it produces a more coherent result because those choices covary in the training data.
Choosing between the three
| What you want | The clause that carries it | Most common failure |
|---|---|---|
| A subject placed in a constructed scene | Chained spatial anchors | Flat composition, elements silently dropped |
| One element restyled in place | An explicit freeze clause | The whole image converts |
| Several described people in one frame | One sentence per subject | Attributes leak between subjects |
Limits worth knowing before you build on it
- Identity drifts across passes, as with any model in this class. Two decisive edits beat six cautious ones.
- Text inside images is inconsistent. Short labels sometimes survive; body copy does not.
- Outputs carry SynthID watermarking. Google applies it to images from its generative models, which matters if provenance detection is part of your delivery process.
- It is a hosted model. No weights to download, no offline run, and availability and terms are set by the provider. That rules it out as a drop-in for a local pipeline.
- The prompts are not portable. These sentence-shaped instructions perform worse on tag-trained diffusion checkpoints, which want condensed noun phrases instead.
Practical takeaway
Write sentences, not tags. Anchor every new element to something already in the frame rather than describing the scene in parallel. Add a literal freeze clause whenever an edit should stay local, and give each person in a multi-subject prompt their own sentence for wardrobe and their own sentence for hair. Rank your requirements explicitly, using soft phrasing for the ones you can live without, because when two clauses conflict something gets dropped, and you want to be the one choosing what.
Top comments (0)