Nano banana prompt format

How to write a prompt nano banana draws the way you meant, on 2.1 and Pro, from text or as an edit. The model draws everything it isn't told: a sign you asked for gets words you didn't write, a person gets a pose you didn't choose, a "ballpark" becomes the most famous ballpark it knows. Every detail you leave out is a guess, and most wrong stills come from those guesses. The requests (endpoint, image order, resolution) are in each task's leaf; how to depict someone or somewhere real is guide-depict-real-people-and-places. Before-and-after prompts from a real job are in references/examples.md.

Layout

Parts in this order, joined by blank lines. Keep the style words identical in every prompt of a job, so the stills match.

Image 1 is <what it is>: use it ONLY for <what to take from it>.
Image 2 is <what it is>: use it ONLY for <what to take from it>.

<The task, one sentence: "Create a new illustration in exactly the art style of image 1." / "Edit image 1: …" / "Redraw image 1 as …">

<The style in words: medium, line, texture, palette.>

<The scene: the camera, the place, each person, the lettering.>

<For an edit: what stays, listed.>

<Aspect ratio.> No other text.

Images

• One line per image, numbered in image_urls order, saying what it is and what to take from it: the art style, a face and hair, the place's architecture, the exact view to keep, the people to add. Say ONLY: the model copies everything else it sees.
• A style reference passes on its content too. A style frame that showed an invented broadcaster and free-standing light towers put that man and those towers into every scene that used it as image 1, and the towers kept coming back in later edits until a prompt said "this ballpark has no light towers; replace them with dark night sky".
• A photo for likeness: "a photo of the broadcaster: use it ONLY for his face and hair", then the clothes, age and pose in words.

Camera and place

• Say where the camera stands and what it faces, by the place's own landmarks: "from behind home plate looking out to centre field", "from the outfield seats looking in toward home plate", "seen over his shoulder, through the booth window". Without it, fans faced a direction that made no sense against the field, and a diamond came back with home plate at the wrong end.
• Name a real place exactly (its name, its former name, the city), and say what is there and what isn't: "the lights hang from the roof trusses; there are no free-standing light towers", "no city skyline". "A classic ballpark … a city skyline behind" got the most famous park of that kind and another city's skyline.
• Time of day, season and era, when the place changed over the years.
• Scale for small things: "a small, realistically sized white baseball" (an unscaled one came back far too big for the people catching it).

People: which way everything points

For each person: where they are in the frame, which way the torso faces, where the head and eyes point (at what), what each hand does and which side of the hand faces the camera, and the posture. Point everything at landmarks in the scene or at the camera.

• "Leaning into a microphone, seen at three-quarter", then "facing out toward the field", came back twisted 90° to the desk with an elbow on the mic: they said where he looked, not where his chest and arms went. "Seen over his left shoulder … chest and shoulders square to the desk, hunched forward with both forearms on it … the microphone between him and the window" landed.
• A fielder seen from behind came back with the glove's pocket toward the camera, the hand turned 180° at the wrist. "The BACK of his glove toward us" fixed the hand; the glove then closed top to bottom until the prompt said "clapped shut lengthwise around the ball, thumb and pinky sides pressed together, fingers pointing up".
• Several people: give each their own sentence, left to right.
• Head and neck: "his head sits naturally on his neck and shoulders, chin level" went in with the move from GPT Image 2.5 to nano banana 2.1, and the unnatural head and hair were gone on both retests; which change did it is untested.

Words in the picture

Current nano banana models draw lettering cleanly. Put words in when they belong to the world (a sign, a banner, a jersey, a scoreboard), and spell each one:

• In double quotes, with where it is: the sign above the video board reads exactly "ACME FIELD". Then "No other text." at the end.
• Anything you ask for that carries words, and don't give words to, gets invented ones: jerseys "with plain blank number patches" came back with garbled numbers and a 0 between a 7 and a 14. Give the numbers, or leave the jerseys facing away.
• Don't ban all text. "No text, lettering, signage or logos anywhere" blacked out the venue's name on its own facade and took away what tells a viewer where they are. Background text no one reads (an exit sign) can be left to the model.
• Real details from a photo: "the retired-number discs keep their numbers"; everything else on it: "replace every advertisement, sponsor logo and piece of text with plain panels, except …".
• A screen with nothing to say: "shows a plain abstract graphic in the palette", or a picture given for it (nano-banana-stage-cast-in-set).
• Titles, captions, lower thirds and chyrons never go in the image: they are composited over it (guide-add-titles-and-motion-graphics).

Edits

• Say what changes, then list what stays: "Keep everything in the ballpark EXACTLY as image 1 draws it: camera position, perspective, ." or "Change nothing else: same building, roof, fans, colours, linework and grain."
• Name the thing the model tends to add back: "do not add any roof structure over the outfield", "do not add a city skyline".
• A removal says what goes in its place ("replace them with night sky and more of the same trees").
• Two fixes in one pass: a numbered list. Three or more: separate passes.

Check it

Before sending, read the prompt for:

• a label line for every image, each saying ONLY what to take from it;
• every real person and place named exactly, and what isn't there said too;
• the camera's position and direction;
• for every person: torso, head and eyes, each hand, posture;
• every word that should appear, quoted, and "No other text.";
• for an edit, what stays, listed.

After the take, check it against the same list and against the reference photos before anyone sees it: a consistent set of stills can still all be of the wrong place.