Face Swap Prompt: 6 That Work, and Why Yours Failed

Writing a Face Swap Prompt That Actually Works

A face swap prompt is not one instruction, it is four: which face changes, where the new face comes from, what must stay exactly as it was, and how the two should blend. Leave out the third and the model helpfully redraws your background too.

But there is a bigger reason prompts fail, and no wording fixes it. Most image models cannot accept a second photo at all. If you are pasting a carefully written face swap prompt into a text-to-image model, it has nowhere to put the face you want to use — so it invents one. That single fact explains the majority of “not working” reports.

This guide covers both: the model check that has to come first, then six prompts you can copy directly, per-model wording, and a fault table for when the output is not what you asked for. Everything runs on Bananaify's face swap tool.

Check the Model Before You Blame the Prompt

A face swap needs two inputs: the photo you are editing and the face you are borrowing. Only models that accept image-to-image input can do this. On Bananaify that is Seedream 5.0 Lite, Flux 2 Pro, GPT Image 2 and Nano Banana 2 for stills, and Seedance 2 or Seedance 2 Mini for footage.

Nano Banana v1, Grok Imagine and Z-Image are text-to-image only. They will happily accept a face swap prompt and return a picture — but it will be a newly generated person, not your face moved into your photo. This is why “nano banana face swap not working” is such a common search: the prompt was fine, the model was wrong. Nano Banana 2 is a different model from Nano Banana, and it is the one that swaps.

Among the four that do work, Seedream 5.0 Lite gives the sharpest identity transfer and is the default in the Face Swap tab. Flux 2 Pro is the next pick. GPT Image 2 handles “change only this person” targeting most reliably, and Nano Banana 2 is often better at matching difficult lighting.

The Four Parts of a Face Swap Prompt

1. Target — who changes

“Transfer only the facial identity onto the person in the first image” is enough for a solo portrait. For a group shot, name the person instead: “onto the woman on the left.” Avoid the obvious phrasing, “replace the face” — these models read face as the whole facial region, and hand you the reference photo's expression and head angle along with it.

2. Source — which features

Name the features rather than the face: “copy from the second image just the identity-defining features: eye shape and colour, eyebrows, nose, lips, and the underlying bone structure.” Be explicit about ordering too, because the model has no other way to know which upload is which.

3. Lock — what must not change

The clause most people leave out, and the one that separates a clean edit from a regenerated picture. It needs to cover the face as well as the frame: “take everything else from the first image and leave it untouched: head angle, gaze direction, facial expression, mouth position, body pose, framing, background, clothing and lighting.” Listing pose alone is not enough — the model reads that as body pose and lets the expression drift.

Pair it with the negative half, which is what actually stops the reference image bleeding through: “the second image is a reference for identity only — ignore its expression, head tilt, gaze, hairstyle, lighting, background and crop.” Saying what to keep and never saying what to discard leaves the model to decide, and it decides badly.

4. Blend — how to merge

“Match the skin tone, lighting direction and colour temperature of the first image. Blend the edges seamlessly so the result looks like an unedited photograph.” This is what stops the result looking like a cutout taped on.

Put together, the default face swap prompt reads:

Face Swap Prompts for Specific Jobs

Keeping the original hairstyle

By default some of the source hair tends to come across with the face, which reads as wrong when the hairline does not match the head shape. Add one clause:

Into a painting or artwork

The trick here is asking for the face to be re-rendered in the medium rather than pasted photographically. Public-domain paintings are the safest source material, since no living person's likeness is involved.

Professional headshots

For team pages where photos were shot years apart, swap current faces onto one consistent set of portraits. Add a sharpening clause, because headshots are viewed large:

Video

A video face swap prompt has to pin down far more, because the model is regenerating the clip rather than editing a single frame. Everything that must survive needs naming explicitly:

The AI face swap video guide covers the per-second costs and the 15-second clip limit in detail.

Why Your Face Swap Prompt Is Not Working

It generated a completely new image

The model is text-to-image. No face swap prompt can fix this — switch to Seedream 5.0 Lite, Flux 2 Pro, GPT Image 2 or Nano Banana 2. This is by far the most common cause.

It changed the background and clothes as well

The Lock clause is missing. Add “keep the original pose, framing, background and clothing unchanged” and the edit narrows to the face.

It swapped the wrong person

The target was described by role, name or counting order — none of which are visible in a photograph. Use position plus one visible detail: “the man on the right in the grey jacket.” There is more on this in our multiple face swap guide.

The result looks flat or pasted on

The Blend clause is missing, or the two photos were lit very differently. Add the skin tone and colour temperature instruction, and where you can, pick a source face shot in comparable light.

The request was refused

Naming a real public figure will get a face swap prompt rejected, as will anything sexual involving a real person. This is deliberate and not something to work around.

Prompts for Gemini and ChatGPT

People search for a face swap prompt for Gemini or ChatGPT because those are the assistants they already have open. Both can edit images, and the same four-part structure applies — attach both photos, then paste the prompt.

Two practical differences are worth knowing. General assistants apply stricter likeness policies, so requests involving a second person's face are refused more often even when consensual. And they give you no control over resolution or output format, which matters if the result is going anywhere real.

Bananaify runs Seedream 5.0 Lite, Flux 2 Pro, GPT Image 2 and Nano Banana 2 directly, which is the same class of model without the assistant layer in between — you pick the resolution, and output comes back with no watermark at up to 4K.

Or Skip the Prompt Entirely

Worth mentioning, since it makes this whole page optional for most people: the Face Swap tab on Bananaify composes the prompt for you from a handful of toggles — keep original hairstyle, match skin tone and lighting, enhance face detail, and a field for naming which person to change in a group shot.

The generated instruction is the same four-part structure described above, and you can expand a preview to read exactly what is being sent before you spend anything. Writing your own face swap prompt is worth it when you want something the toggles do not cover — a painting, a specific artistic medium, or stacked edits like swapping the face and changing the outfit in one pass.

Whose Face You Can Use

Your own face, someone who agreed, or a public-domain artwork is fine. Using a face swap prompt to impersonate a real person, fabricate something that looks like evidence, or produce sexual imagery of someone violates Bananaify's terms and, in a growing number of jurisdictions, the law. Explicit and non-consensual requests are blocked, and prompts naming public figures are refused.

Uploads are used only to produce your result — not sold, not used to train models, not published anywhere — and generated files can be deleted from the dashboard at any time.

Face Swap Prompt FAQ

What is the best face swap prompt?

The four-part structure: name the target, name the source, lock what must not change, then describe the blend. The default version is in the section above and works for most photos without modification.

Why is Nano Banana face swap not working?

Nano Banana v1 is text-to-image only — it cannot accept your second photo, so it generates a new face instead of swapping yours in. Use Nano Banana 2 or GPT Image 2, both of which take two images.

Can I use a face swap prompt in Gemini or ChatGPT?

Yes, with both images attached. Expect more refusals on likeness grounds and no control over output resolution. The same four-part wording applies.

How do I stop it changing the background?

Add the Lock clause: “keep the original pose, framing, background and clothing unchanged.” This is the single most valuable line in any face swap prompt.

Which model gives the best results?

Seedream 5.0 Lite first, then Flux 2 Pro, GPT Image 2 and Nano Banana 2. All four are available in the Face Swap tab; a photo swap costs 4 credits at standard resolution, and the exact price for the model and resolution you picked is shown on the button before you generate.

Do I need to write a prompt at all?

No — the Face Swap tab builds one from toggles. Write your own when you want a painting style, a specific medium, or several edits stacked in one pass.

Start Swapping

Check the model first, then use the four-part structure: target, source, lock, blend. Nearly every failed swap traces back to one of those two things.

Open the face swap tool to run any prompt on this page. For related reading, see multiple face swap for group photos, swapping faces in video, and our wider AI photo editing prompts collection.