Gemini Image Generation Prompts: How to Write Them

Gemini (gemini.google.com) makes images inside an ordinary chat. There is no prompt box with sliders, no seed field and no style menu. You describe what you want in a sentence, Gemini generates it with Google’s image models, and you keep talking to change it. That conversational design is the whole point, and it changes how you should write. Keyword soup that works in older generators tends to underperform here; a clear brief written for a human art director tends to work very well.

How image generation works in Gemini

Gemini's image generation and editing run on Google's Nano Banana models. At the time of writing, Nano Banana 2 is available for free and Nano Banana 2.1, which Google calls its best image generation and editing model, comes with a Google AI plan. Both handle new images and conversational edits, keep characters consistent, render text and combine photos. Google's older Imagen models are now legacy. Model names, limits and resolution differ by plan and region.

Every image Gemini produces carries a SynthID watermark, an invisible marker that identifies it as AI-generated. It does not affect how the image looks, but it is worth knowing if you deliver work to clients. Availability, limits and resolution vary by plan and region, so check the current Gemini help pages.

Because everything runs in a chat, your history is the only record of what you asked. Give each image project its own chat and name it.

The anatomy of a good Gemini prompt

Write a brief, not a tag list. A reliable order is subject, action or pose, setting, lighting, style and format. Each part answers a question the model would otherwise guess at.

• Subject. Be specific. "A grey tabby cat" beats "a cat". "A 1970s orange moped with chrome mirrors" beats "a scooter".
• Action or pose. "Curled asleep", "mid-stride", "held up to the light". Verbs give the image life and stop the stiff, centred default.
• Setting. Where and when. "On a tiled Lisbon doorstep at dusk" places the subject and implies colour and shadow.
• Lighting. This is the single biggest quality lever. "Soft overcast daylight", "hard noon sun", "warm tungsten lamp from the left", "backlit by a window".
• Style and medium. "Photograph, 50mm lens", "flat vector illustration", "loose watercolour", "3D render with matte materials". Name the medium rather than a living artist.
• Format. Ask for the orientation you need: "square", "portrait 4:5", "wide 16:9". Aspect ratio control varies by plan, so say it in words and check the result.

Put it together: "Create a photograph of a grey tabby cat curled asleep on a tiled Lisbon doorstep at dusk, warm light spilling from an open door, 50mm lens, portrait orientation." One sentence a human could brief from: that is the test.

Refining in the conversation

The second message matters as much as the first. Instead of rewriting the whole prompt, tell Gemini what to change and what to keep: "Keep everything the same but make the light cooler and the cat awake, looking at the camera." Treat it like directing a photographer between shots.

1. Generate once with a full brief. A good first image gives you something to steer.
2. Change one thing per message. Light, then pose, then background, so you know which instruction caused any drift.
3. Name what must stay. "Same cat, same doorstep" anchors the subject when you ask for a new angle.
4. Ask for variations explicitly. "Give me three options with different colour palettes" gets you choice without a new chat.
5. Copy the final prompt out. Paste the full chain of instructions into a note alongside the saved image. Chats are easy to lose.

When you want to edit an existing photo rather than generate from nothing, upload it and describe the change; the Gemini image editing guide covers that workflow in detail.

Settings at a glance

Gemini hides most controls behind language, so the "settings" are phrases you include in the prompt.

Common problems and fixes

• The image is generic or stock-like. Add a specific place, a time of day and a small imperfection ("chipped paint", "a dropped glove"). Specific detail is what makes an image feel seen rather than assembled.
• Text comes out misspelt. Keep lettering to a few characters, put it in quotes, and say where it goes ("the word SALE on a hanging card"). For longer text, leave space and add it in a design tool.
• The subject changes between messages. Say "same subject" and describe it again briefly, or edit an uploaded image instead of regenerating.
• Gemini declines the request. Prompts involving real people, brands or sensitive content may be refused. Describe a fictional person and an unbranded product instead.
• Wrong orientation. Restate it at the end of the prompt and check whether your plan supports the ratio.

Try it

Create a photograph of a small ceramic teapot with a matte sage glaze on a crumpled linen cloth, a single steaming cup beside it, soft overcast window light from the left, 50mm lens, shallow depth of field, muted natural colours, square format.

Run it once, then send "Keep everything the same but switch to warm late-afternoon sun and add a sprig of mint on the saucer". Compare the two to see how much the lighting phrase alone changes.