We ran a one-line nano banana face swap prompt and a structured one through Nano Banana 2 back to back, same two reference images. swap the face brought the source hairstyle across with the face; the structured face swap prompt did not. What separated the two was a short list of what must not change.
Below is the prompt that fixed it, the four slots it fills, and the side-by-side result from our own runs.
Last updated: July 2026
TL;DR
- A one-line face swap prompt produces a person swap, not a face swap. Same model, same inputs — the source hairstyle came along and overwrote the target's.
- Name what to keep. The clause that changed our result was an explicit keep-list: pose, hairline and hair, clothing, background, framing.
- Scope the identity with "only". Take only bone structure, eyes, nose, mouth, stubble, skin texture from the source.
- Restate the target's light. A studio face dropped into dusk reads as fake until you name the direction, softness and colour temperature.
- Cost: each run is 8 credits on Nano Banana 2 at 1K. Our whole test — two inputs plus two swaps — was 32 credits.
Why a one-line face swap prompt is not enough
We ran the naive face swap prompt and the structured one through the same model, with the same two reference images, minutes apart.
Input 1 was a studio headshot: flat frontal key light, short hair, grey backdrop. Input 2 was a figure on a rainy street at dusk — long wet hair, three-quarter turn, cool teal grade.
The one-line face swap prompt did transfer the identity, and it matched the scene light better than expected. But it also brought the source hairstyle with it. The output is a different person standing in the target's coat, not the target wearing a new face. If your goal was to keep a character consistent across shots, that result is unusable.
The structured face swap prompt held the target's wet long hair, the head turn, the coat, the grade and the light direction, and changed only the facial identity. Same model, same two inputs, minutes apart — the face swap prompt was the only thing that differed.
We cannot tell you why from two runs — that would need a lot more than four images. What the pair does establish is narrower and still useful: on this model, with these inputs, the word "face" alone did not confine the edit to the face, and naming the hair explicitly did. If you are relying on a one-line prompt to hold a character steady across shots, that is worth an 8-credit test of your own before you build on it.
The four slots a face swap prompt has to fill
1. KEEP — what must not change. This is the clause our two face swap prompts differed on, and hair was what it saved: named, the target kept its wet long hair; unnamed, the source's short hair came across.
2. TAKE — scope the identity narrowly. We wrote "take from the first image only: bone structure, eye shape and colour, nose, mouth, stubble pattern, skin texture", and the target's head turn survived rather than snapping to the source's straight-to-camera pose.
3. MATCH — restate the target's light in words even though it is visible in the image. Ours read "cool blue-teal dusk from the upper right, soft, no hard shadow", plus a request to reproduce the grade and grain on the new face. The source was flat frontal studio light, so this was the widest gap in the test.
4. BLEND — we asked against two artefacts in one sentence: a seam or halo at the jaw and hairline, and smoothed-over skin ("photorealistic, no smoothing, keep pores visible").
One structural choice in our face swap prompt worth copying: we passed the images in a fixed order — source face first, target scene second — and referred to them as "the first image" and "the second image" in the prompt. That removes any question about which image is being edited; we did not test what happens if you leave it ambiguous.
The face swap prompt template
Copy this face swap prompt and replace the bracketed parts. This is the exact structure from the run above:
Replace the face of the person in the second image with the face of the person in the first image.
Keep from the second image: the exact head pose and [angle], the hairline and hair, the [clothing], the body position, the [background], and the framing.
Take from the first image only: facial identity — bone structure, eye shape and colour, nose, mouth, [facial hair], and skin texture.
Match the second image's lighting: [colour and direction of the light], [soft or hard], and reproduce its colour grade and film grain on the new face so the skin tone reads as the same exposure.
Blend the jawline and hairline edges so there is no seam or halo. Photorealistic, no smoothing, keep pores visible.
Fill every bracket with something specific. "Cool blue-teal dusk from the upper right, soft" is the level of detail that was in the run above; we have not tested how much of that specificity you can drop before the result changes.
When the face swap prompt still misses
Four ways a face swap prompt still comes back wrong, and what actually moves them.
The face looks pasted on. The MATCH slot is too vague. Describe where the light comes from and how hard it is — "cool blue-teal dusk from the upper right, soft, no hard shadow" beats "match the lighting".
Identity drifted. Our source was a head-and-shoulders crop and the gap it had to cross was flat frontal studio light to a three-quarter turn at dusk. The structured prompt held across that; we have not measured how much further it stretches before it breaks, so treat a harder pose gap as untested rather than as solved.
Skin went plastic. Add the texture clause to BLEND and keep it in the same sentence as the seam request. Asking for photorealism alone is not enough.
Hair changed anyway. Check that hair is actually in the KEEP list. We named it as "the hairline and hair" — two things, because a hairline can be redrawn without the length changing and the reverse is also true.
Regeneration is cheap enough to iterate honestly: one more 1K attempt costs 8 credits, under half the 18 a single 4K render takes.
What a face swap prompt run costs
A face swap prompt run on Nano Banana 2 image-to-image is 8 credits at 1K on Kavel — 12 at 2K, 18 at 4K. Our entire test, two reference images plus two face swap attempts, came to 32 credits. The cost preview shows before you start and failed jobs are not charged, so a bad face swap prompt costs you a retry, not a balance.
For context on the credit pool: Starter is $248/year for 33,500 credits, which is a little over 4,000 runs at this size. You can run the model directly on the Nano Banana 2 page, and the same prompt structure works on the AI gender swap and AI clothes swap tools, which are the same editing pipeline with a different keep-list.
FAQ
What is the best nano banana face swap prompt?
The one that names what to keep. In our test, swap the face replaced the hair along with the face; a prompt with an explicit keep-list — pose, hairline and hair, clothing, background, framing — changed only the facial identity. The full template is above.
Why does my face swap prompt swap the hair as well as the face?
In our test it did so whenever the hair was not named. We are not claiming to know the mechanism from two runs — only that adding "the hairline and hair" to an explicit keep-list was the change that stopped it, with everything else held constant.
Does the order of the two images matter?
Yes. Put the source face first and the target scene second, then refer to them as "the first image" and "the second image". Ambiguity about which image is being edited is a common cause of a swap running backwards.
How many credits does a face swap cost?
8 credits per run at 1K on Nano Banana 2. Failed jobs are not charged, and the cost preview appears before you start the run.
Can I use the same prompt structure for other edits?
Probably, but we tested this face swap prompt structure on faces only. The AI gender swap and AI clothes swap tools run the same image-to-image pipeline on the same model, so the same four clauses are the obvious thing to try there — with the keep-list adjusted for what you want held.
Will the same face swap prompt give the same result twice?
Not exactly. These models are not deterministic, so expect variation between runs even with identical inputs. Judge a face swap prompt on how often it lands, not on one attempt.
Sources
- Our own runs on nano-banana-2 image-to-image via the model endpoint this site uses, July 2026 — 4 runs, 8 credits each, reference images are AI-generated fictional people
- Kie provider price sheet for nano-banana-2 (1K / 2K / 4K tiers), read July 2026
- Kavel pricing and live model configuration, read from the site in July 2026
