A headshot asks one thing of an image model that a landscape never does. The person has to still be the person. Google publishes both the mechanism and the failure mode, so most of the work becomes deciding what the prompt locks and what it leaves free.
Nano Banana Pro headshot avatar: the short answer
Upload one clean reference, then write two clauses. The first holds the identity. The second changes the styling.
Google’s own wording does both in a single sentence: the man “should retain his original features and expression”, while harsh directional light is introduced across the face.
Write only the styling half and the model is free to reinterpret the face. That is the drift people describe as a flattering stranger.
What Google documents about holding a face
DeepMind’s model page publishes the ceiling. The model can “maintain the consistency and resemblance of up to five characters and the fidelity of up to fourteen objects in a single workflow”. Those are upper bounds for one workflow, not a promise about your photo.
The limitations section on the same page is just as direct: “The model excels at character consistency, but it may not always get it right.” Small faces get a line of their own, because the model “can still struggle with small faces, accurate spelling, and fine details in images”. The prompting guide repeats the caveat for edits: character consistency across edits “may vary”.
Read the three together and the working rule appears. Consistency is a property of a session, not of the model. Keep one reference and one wardrobe inside one session and the odds hold up.
Give every uploaded image a job
The prompting guide carries a rule for references: “clearly define the role of each”. Its own example names three files and three jobs, one for pose, one for art style, one for the background.
A headshot session usually needs two. The subject photo carries the face. A second image carries the light or the background, and you say so out loud.
Write the identity lock first
Separate the two lists before writing anything.
What changes: wardrobe, background, light direction, colour grade, framing, aspect ratio. The model page lists all of those among the supported moves.
What stays: the face, the hairstyle, glasses when they are part of the look, and any marker that makes the person recognisable.
Google’s locking clauses are short and repetitive on purpose. One reads “keep the can exactly the same”. Another reads “The character, remains exactly locked in its current position”.
One warning about the internet’s favourite magic words. “Preserve face 100%” is not Google’s phrasing and does not appear in the documentation. Copy the plain shape the model page publishes instead.
Worked examples: three headshots
Three prompts, each built on a sentence Google publishes, with the reference roles filled in. The added lines are flagged under each block.
Example one: relighting an existing portrait
Starting point: one well-exposed portrait, plain background, the face readable at full size.
Use the uploaded portrait as the only identity reference. Keep the face, the hairstyle and the glasses unchanged. Generate an image with an intense chiaroscuro effect. The man should retain his original features and expression. Introduce harsh, directional light, appearing to come from above and slightly to the left, casting deep, defined shadows across the face. Only slivers of light illuminating his eyes and cheekbones, the rest of the face is in deep shadowThe chiaroscuro passage is Google’s, verbatim. The opening line is the reference role the prompting guide asks for. Expect a darker key on the same face, and check the eyes at thumbnail size, because that is where drift shows first.
Example two: one reference, three crops
Starting point: the same reference, needed at three framings for a resume, a profile picture and a channel banner.
Keep the uploaded reference as the exact same person in all three images. change aspect ratio to 1:1 by reducing background. The character, remains exactly locked in its current positionThe lock sentence is Google’s, verbatim. The model page lists nine ratios in its own examples, among them 1:1, 4:3, 5:3, 1.85:1, 2.39:1, 2.75:1, 4:1, 9:16 and 1:4. Expect the same subject with the background trimmed rather than the face redrawn. Google does not publish a full supported list.
Example three: one team photo, six people
Starting point: six separate portraits for a team page, shot in different rooms on different days. The second half is Google’s, verbatim; swap the fashion-editorial phrase for the look your team wants.
Use the six uploads as identity references, one person per image, no mixing of features. Put these six people into a single image, they should fit into a stunning award-winning shot in the style of a fashion editorial. the identity of all six people and their attire must stay consistent throughout but they can and should be seen from different angles and distances in as is most natural and suitable to the sceneThe documented ceiling is five characters per workflow, so six is already past it. Expect one frame with six readable identities, the harder version of the same problem. Split the group and composite, or accept a re-roll.
The square avatar is a different problem
A 1:1 crop removes margin, which is where a face loses the top of the head or the point of the chin. Ask for the ratio change and the lock in the same breath, as in the second example.
Safe area is not a setting the documentation exposes. If a platform crops to a circle, keep the face centred and ask for room around it in the prompt. Then look at the result at the size people will see it. A 4K headshot that turns to mush at 40 pixels wide has failed the job it was made for.
For the wider framework behind these prompts, the Nano Banana Pro prompt guide covers the rest, and the how-to guide shows where the reference upload lives. The photo editor article handles images you already have, and the Nano Banana Pro vs Nano Banana comparison explains which generation to reach for. Builders should continue with the API guide.
Frequently asked questions
Will the headshot look like me? Often, not always. Google writes that the model “excels at character consistency, but it may not always get it right”.
How many reference photos should I upload? Up to fourteen images in one workflow, with a ceiling of five characters. The role sentence matters more than the count.
Can one reference give me a square avatar and a portrait crop? Yes. Ask for the ratio change and add Google’s locking clause in the same instruction.
Should I write “preserve face 100%”? No. That wording is not in the documentation. Use the plain locking clause the model page publishes.
Look at the eyes at 40 pixels wide before you ship the studio version.