Nano Banana Pro Person Swap Prompt: Face Swap, Full Body, Consent

A Nano Banana Pro face swap works when the prompt gives each photo one job: Image 1 supplies the identity, Image 2 supplies the scene, and one sentence says that everything else in Image 2 stays exactly as it is. That lock sentence comes from Google's own editing template. The same prompt works with Nano Banana 2, which is what the Gemini app uses by default.

Use it on photos you have the right to edit: yourself, a friend who said yes, a model with a signed release, or a character you made. Google's systems look for "using someone's image without their consent," so consent is the first step of this workflow.

Copy this base prompt:

Image 1 is the identity source. Image 2 is the target photo.
Replace the person in Image 2 with the identity of the person in Image 1.
Keep everything else in Image 2 exactly the same, preserving the original pose, clothing, lighting, camera angle, background, and composition.
Match skin texture, shadows, perspective, and white balance so the result looks like one photo.

Then pick a branch and add its line:

The five parts of a person swap prompt

A weak request such as "swap this person into this photo" leaves the model to guess five things: which image owns the face, which owns the scene, how much should move, what must not change, and what result you would reject. Each guess is a chance for the wrong face, a new background, or a borrowed outfit. A reliable swap prompt answers all five in plain words.

The locked-attributes sentence does most of the work. Google's image-generation guide gives this template for targeted edits: "Using the provided image, change only the [specific element] to [new element/description]. Keep everything else in the image exactly the same, preserving the original style, lighting, and composition." A face swap is a targeted edit where the new element comes from a second image, so the base prompt above follows the same pattern.

Phrase the avoid list as concrete, visible changes. Google's guide recommends describing the scene you want rather than stacking negatives, so "keep the street empty behind her" works better than "no extra people." A short avoid line for the mistakes you have already seen is fine. A long list of forbidden words is not.

Prepare Image 1 and Image 2 before you write more words

Better reference photos fix more swaps than extra adjectives. Check the two images first.

Image 1, the identity source. Pick a sharp, front-facing or three-quarter photo with even light. Sunglasses, heavy filters, motion blur, a hand across the face, or a steep camera angle force the model to guess bone structure and skin tone. That guess is where "close, but not the same person" results start. If you can, use a source photo with a head angle similar to the target.

Image 2, the target. Decide what it owns. In a face-only swap it owns the body, hair outline, clothing, hands, pose, background, light, and camera. In a full-person swap it owns the scene, camera, light, scale, and the spot where the person stands. If the prompt does not say which, the model may blend the two halfway or rebuild the scene around Image 1.

More than one source photo. Extra photos of the same person can help the model hold the face. Google's API guide sets different caps per model:

When Image 2 is a group photo, name the target by something visible: clothing, position, or what the person is holding. "The man in the gray hoodie, second from the right" is easier for the model to follow than "the person on the right." Then add one sentence that keeps everyone else unchanged.

Face-only swap prompt for Nano Banana Pro and Nano Banana 2

Use the face-only branch when Image 2 already has the pose, outfit, and setting you want. It is the narrowest edit, so it is also the easiest to control.

Image 1 is the identity source. Image 2 is the target photo.
Change only the face and facial features of the person in Image 2 so they match the person in Image 1.
Keep the body, hair outline, clothing, hands, pose, background, camera angle, lighting direction, depth of field, and composition of Image 2 exactly the same.
Match skin texture, facial shadows, perspective, white balance, and film grain so the face belongs in Image 2.
Do not change any other person or object in Image 2.

The key phrase is "change only the face and facial features." Without that limit, the model may borrow hair, clothing, or body shape from Image 1. That is useful in a full-person swap and wrong everywhere else.

This branch suits headshots, trying a new hairstyle or outfit on yourself, putting yourself into a travel or event photo, and restoring family photos with the family's agreement. Google's Gemini Apps help describes Nano Banana 2 as "ideal to create personalized images such as when you want to view yourself with different hairstyles, picture yourself in different scenes."

Full-person swap prompt: who owns the body and the clothes

Use the full-person branch when the person in Image 1 should replace the person in Image 2. It is harder because the model has to decide whose body, clothing, and scale survive. Make that decision for it.

Image 1 is the identity and person reference. Image 2 is the target scene.
Replace the person in Image 2 with the person from Image 1.
Keep the scene, camera position, lighting, background, perspective, and composition of Image 2 exactly the same.
Place the new person in the same spot, at the same scale and depth, with a natural pose for that spot.
Clothing: keep the outfit from Image 1.
Do not change any other person, object, or text in Image 2.

Change the clothing line to fit the job. Write Clothing: keep the outfit worn in Image 2 when the scene's outfit matters, as with a uniform, a costume, or a matching set. If you want the face from Image 1 on the body from Image 2, that is a face-only swap, so use the branch above. Most strange full-person results come from mixing these three jobs in one prompt.

When one pass is not enough, split it. Run the swap first. Then, in the same chat, follow up with one fix at a time: "Keep the person's face exactly the same. Warm the light on the face to match the sunset behind them." Multi-turn editing keeps earlier choices in context, and a one-change follow-up is easier to check than a long prompt that tries to fix everything at once.

Where to run it: Gemini app, Redo with Pro, or the API

The prompt is the same everywhere. What changes is which model answers and what you need to run it, as of October 5, 2026.

In the Gemini app, upload both photos in one message, type the prompt, and submit. Choose the Flash or Pro chat model so the app uses Nano Banana 2, which accepts multiple reference images. Google's help page says Nano Banana 2 Lite "isn't designed for multiple image references." Redo with Pro also stops once you reach your daily Nano Banana 2 image quota.

Since Nano Banana 2 launched on February 26, 2026, it has replaced Nano Banana Pro across the app's Fast, Thinking, and Pro models. Pro remains as the redo option for plan subscribers, so a free account no longer gets Nano Banana Pro in the Gemini app. For the remaining no-cost options, see Can You Use Nano Banana for Free?.

API model IDs. Use gemini-3-pro-image for Nano Banana Pro. The preview IDs gemini-3-pro-image-preview and gemini-3.1-flash-image-preview were shut down on June 25, 2026, so old code that still calls them fails.

API cost per swap, at Google's Standard rates as of October 4, 2026:

• Nano Banana Pro: $0.134 per 1K or 2K image and $0.24 per 4K image, plus about $0.0011 per input reference image. Thinking is on by default and billed as output tokens.
• Nano Banana 2: $0.067 per 1K image and $0.101 per 2K image.
• Batch requests cost half of Standard on both models.

The snippet below sends two images to Nano Banana Pro through the Interactions API. It follows the request shape in Google's image-generation guide (updated September 23, 2026) and has not been executed. Try it with two photos of yourself before using anyone else's.

import base64
import os

from google import genai

client = genai.Client(api_key=os.environ["GEMINI_API_KEY"])


def image_part(path, mime_type="image/jpeg"):
    with open(path, "rb") as f:
        data = base64.b64encode(f.read()).decode("utf-8")
    return {"type": "image", "mime_type": mime_type, "data": data}


prompt = (
    "Image 1 is the identity source. Image 2 is the target photo. "
    "Change only the face and facial features of the person in Image 2 "
    "so they match the person in Image 1. Keep everything else in Image 2 "
    "exactly the same, preserving the original pose, clothing, lighting, "
    "background, and composition."
)

interaction = client.interactions.create(
    model="gemini-3-pro-image",
    input=[
        {"type": "text", "text": prompt},
        image_part("identity.jpg"),  # Image 1: the identity source
        image_part("target.jpg"),    # Image 2: the target photo
    ],
    response_format={"type": "image", "image_size": "2K"},
)

with open("swap.png", "wb") as f:
    f.write(base64.b64decode(interaction.output_image.data))

Send the identity photo first so the order matches "Image 1" and "Image 2" in the prompt. By default the output takes the size of the input image. Add an aspect_ratio field to response_format only if you need a different frame. For JSON request templates and the response shape, see the Nano Banana Pro API guide.

Consent comes first: what Google detects and enforces

A person swap changes how a real person appears, so the question of whether you may make it comes before the prompt. No wording makes a non-consensual swap acceptable. A clean prompt makes a consented one work better.

Go ahead when the face is yours, when the person in Image 1 agreed to this specific use, when you have a model release that covers AI edits, or when the character is fictional and you made it. Gemini also offers an avatar feature: you can record your own face and voice and use that avatar when you create images with Nano Banana. It is the most direct consented route for putting yourself into new scenes.

Stop if any of these is true:

• You do not have the person's permission. Google's Gemini Apps enforcement page lists "using someone's image without their consent and non-consensual intimate imagery" among the misuse its automated systems and human reviewers detect.
• The result would be sexual or intimate. Non-consensual intimate imagery is prohibited outright under Google's Generative AI Prohibited Use Policy.
• The result could pass as real and deceive someone: a fake event, fake evidence, a fake endorsement, or a scam. The same policy prohibits impersonating a person without disclosure to deceive, and using personal data or biometrics without legally required consent.
• The swap involves a child. Child-safety protections are always on in Google's models, and on the API they cannot be adjusted.

What enforcement looks like. In the Gemini app, images can be removed when Google's systems detect a possible violation. Confirmed violations bring in-product or email notices. Repeated violations can restrict your use of Google's generative AI products or your Google account, and the notice includes an appeal link. On the Gemini API and AI Studio, Google's usage-policies page describes email contact, temporary usage limits, temporary suspension, and as a last resort closing access to the Gemini API and other Google services. Suspension and closure notices include an appeal link. API prompts and outputs are kept for 55 days for abuse detection.

Rewording a refused swap is its own violation. The Prohibited Use Policy lists "circumvention of abuse protections or safety filters" as a prohibited use, and the Gemini enforcement page names attempts to circumvent safety protections as something it detects. If the model refuses a swap of someone else's photo, treat the refusal as the answer.

Disclose the edit when you share it. Every image from these models carries an invisible SynthID watermark. In the Gemini app, a visible watermark is a setting under Settings > Media Watermark. Google notes that users in India, South Korea, and Vietnam see that setting only with an AI Ultra subscription and otherwise always get the visible mark. Turning off the visible mark does not remove SynthID or Content Credentials, and anyone can upload an image to the Gemini app and ask whether Google AI made it. When a swapped photo could make a viewer believe something happened that did not, say it is edited. If it cannot be shared honestly with that label, do not share it.

Fix the wrong person, a pasted face, background drift, or pose drift

When the result is wrong, find the failure type before you add style words. Most problems come from an unclear role, not from a missing adjective.

If two failures show up at once, fix the role first and the realism second. A pasted-looking face can usually be repaired with a blending line. A face on the wrong person means the model never understood which person was the target.

When nothing comes back. On the API, a blocked edit returns no image, and generateContent responses report a reason such as IMAGE_SAFETY, IMAGE_PROHIBITED_CONTENT, IMAGE_OTHER, or a promptFeedback.blockReason. If you are editing your own photo and still get a refusal, the cause may be an age or account condition, a quota, or a false positive. The checks in Why Can't Gemini Generate Images of People? separate those cases. Nano Banana 2 blockReason OTHER covers the vague "OTHER" result, and Nano Banana Pro Safety Filters explains which filters you can adjust and which stay on. If the photo shows someone who has not agreed to the edit, the refusal is the policy working, and the answer is not to retry.

Check identity, scene, and consent before you share the swap

No prompt guarantees the same face every time. Google lists character consistency as a Nano Banana 2 feature, but user reports of faces drifting after edits are common, and the source photo, the head angle, and the light all affect the result. Check every output before it leaves your device:

• Identity: would someone who knows the person in Image 1 recognize them, or does the face drift toward a stranger?
• Scene: did the pose, body, clothing, background, camera angle, and light follow the branch you chose?
• Blending: do skin texture, edges, shadows, white balance, and grain match the rest of Image 2?
• Side effects: did any other face, sign, logo, or object change?
• Consent and disclosure: has the person agreed to this use, and will viewers know the photo is edited?

Use the checks in this order: consent and disclosure, then identity and scene, then blending. A failed consent check ends the job. A failed scene check means the roles need rewriting. A failed blending check needs only one more follow-up line.

If you only want to replace what is behind a person and leave the person untouched, that is a different edit. The prompts in Gemini Image Background Change handle it with less risk to the face.

FAQ

What is the best prompt for a face swap with Nano Banana Pro?

There is no magic phrase. The prompt that works best names Image 1 as the identity source and Image 2 as the target, says "change only the face and facial features," and ends with "keep everything else in Image 2 exactly the same." Add a blending line for skin, shadows, and white balance, and name the target by clothing and position in a group photo.

How do I face swap in the Gemini app?

Open Gemini signed in to a personal account (you must be 18 or older to edit images), upload the identity photo and the target photo in one message, and paste the face-only prompt. Gemini uses Nano Banana 2 by default. If you have a Google AI Plus, Pro, or Ultra plan, open the result and choose More > Redo with Pro to rerun it with Nano Banana Pro.

Is Nano Banana Pro better than Nano Banana 2 for person swaps?

Nano Banana Pro accepts up to 5 character reference images compared with 4 on Nano Banana 2, and it reasons through dense instructions before drawing, which helps with group photos and full-person swaps. Nano Banana 2 is available to everyone in the Gemini app and costs half as much per 1K image on the API ($0.067 against $0.134). Start with Nano Banana 2 and use Pro for the swaps it gets wrong.

Can I face swap Nano Banana Pro for free?

Partly. The Gemini app's default Nano Banana 2 handles swaps without a plan, within standard usage limits. Nano Banana Pro in the app needs a Google AI plan; the one documented no-plan Pro route is AI Mode in Google Search, for adults in the US, in English. The Gemini API has no free tier for image output, so API swaps are paid from the first image.

Can I swap in a celebrity or someone else's photo?

Only with that person's permission, and never to deceive. Google detects "using someone's image without their consent," and its policy prohibits impersonating a person without disclosure to deceive. A refusal on someone else's face is the policy working, and rewording to get around it is itself a policy violation.

Can a prompt keep exactly the same face every time?

No. A clear prompt and a good identity photo make a close match more likely, but identity can still drift between runs. Rerun with "keep the person's face exactly the same" in a follow-up turn, and review every result before you use it.

Does a swapped image carry a watermark?

Yes. Every image from Google's image models carries an invisible SynthID watermark. The visible Gemini sparkle is a Gemini app setting, which users in India, South Korea, and Vietnam can change only with AI Ultra. Turning it off does not remove SynthID, and anyone can ask the Gemini app whether an image was made with Google AI.