Instructions, not keywords
Gemini’s image models are built on a conversational assistant, so they handle requests the way a person would: “Create a photo of…, make the lighting…, keep the background…”. Keyword lists and weight syntax don’t help here. Our Nano Banana format starts with a verb such as “Create”, then walks through the subject, composition, light, colours, style and mood in short sentences.
Clear instructions also make results more predictable. When each sentence covers one aspect of the image, it’s easy to see which part of the prompt to change when something isn’t right.
It also helps to state the purpose of the image, such as a blog header or a product shot, because Gemini uses that context to choose sensible framing.
Use the prompt as a starting point for edits
Nano Banana is especially good at editing through follow-up messages. Paste the generated prompt to create a first version, then refine it in the same conversation: “make it golden hour”, “move the cup to the right”, “remove the text on the sign”. Short follow-ups work better than rewriting the whole prompt.
You can also give Gemini your original image together with the prompt and ask for a variation. The description then acts as a checklist of what to keep.
Saying what to avoid
Each result ends with a sentence about what to leave out, for example no text, no watermark or no extra people, when that’s relevant to your image. Plain-language exclusions work well with Gemini, unlike models that need a separate negative prompt field.
If you need a particular format, say so in words, for example “in a tall 9:16 format” or “as a square image”. Gemini understands these requests in the prompt itself.
Prompts in your language
Gemini understands many languages, so writing the prompt in your own language works better with Nano Banana than with most image models. Choose from 14 output languages and keep the prompt in the language you think in.
What a result looks like
For a photo of a coffee cup by a window, a result might read: “Create a photo of a speckled ceramic mug on a linen table beside a tall window. Frame it as a close-up from slightly above. Use soft morning light from the left with gentle shadows. Keep the colours warm and neutral, with a calm, unhurried mood. Don’t add any text or logos.” Each sentence covers one aspect, so you can change one without touching the rest.
Guides from our blog
- How to reverse-engineer any image into a prompt, and when it works
A practical reverse prompting workflow: describe an image, format it for your generator, and refine one element at a time. What works, and what never will.
- Why your AI images look generic, and seven prompt habits that fix it
AI image prompts that say how a picture should look, not just what is in it: seven habits, with before-and-after examples, for sharper results.