How do I keep the same face from a photo when generating new AI images?

Upload one or two clear front-facing photos, say which features must not change, then describe the new scene. Data from 2,676 Miku AI tasks.

Published · 5 min read

Short answer: Upload one or two clear, front-facing photos of the same person as reference images, write "keep the facial features, face shape, hairstyle and skin tone of the person in image 1 unchanged", then describe the new scene, outfit and pose. On Miku AI, 233 users ran 2,676 of these tasks in the last 30 days; GPT Image 2.6 Sunburst produced an image in about 79% of 239 tasks, and GPT Image 2.6 Flare in about 95% of 40.

Why keeping a face is hard

When a model reads a reference photo, it takes in the face together with the hair, lighting, expression and background. If the prompt just says "use this person in a beach photo", you may get the right mood and a face that is only similar.

So you need two things: a clear enough face for the model to work from, and words that say which features must match and which may change.

What Miku AI users actually did

We looked at tasks on Miku AI over the last 30 days (up to November 2026) whose prompts asked to keep the same face or the same person:

  • 233 users submitted 2,676 tasks, all with reference images. The median was 2 reference images per task.
  • About 69% produced an image. About 17% were refused by content moderation, much higher than for simple one-spot edits, because almost every reference is a photo of a real person.
  • 16:9 was the most common aspect ratio (1,018 tasks), then 3:4 (607). About 56% chose 1K output.

Results by model (only models with at least 20 tasks on this job; median time for successful tasks):

ModelTasksSuccessMedian timeCredits per 1K image
GPT Image 2.6 Sunburst23979%53 s0.3
GPT Image 2.6 Flare4095%54 s0.3
Nano Banana 27279%63 s0.25
GPT Image 2.5 Flare20474%78 s0.3
GPT Image 21,20667%79 s0.3
GPT Image 2.5 Sunburst80867%85 s0.3
Nano Banana Pro3367%79 s0.35

Start with GPT Image 2.6 Sunburst: a solid sample, around 80% success and under a minute per image. GPT Image 2.6 Flare had the best rate, but on only 40 tasks. Nano Banana 2 also reached about 79% and is the cheapest at 1K. GPT Image 2, the most used model, did worse on this job at about 67%, and most of its failures were moderation refusals (about 19% of its tasks).

Success only means an image came back; whether it looks like the person is your call. For several people in one image, see the people and object limits in Nano Banana Pro's reference image limit.

Step by step in the workspace

  1. Pick one or two photos of the person: front-facing, evenly lit, no sunglasses or mask, with the face reasonably large in the frame.
  2. Upload them as reference images, the portrait first, any scene or outfit images after it.
  3. Choose GPT Image 2.6 Sunburst or Nano Banana 2.
  4. Choose the aspect ratio for where the image will be used; 2K helps with facial detail.
  5. Write the prompt in this order: who, what to keep, new scene, new outfit and pose, image style.
  6. Compare eyes, nose, mouth and face shape with the original. If it is off, try a sharper photo rather than rerunning again and again.

Prompt templates

New setting:

Image 1 is a front-facing photo of a woman. Make an outdoor professional headshot of her. Keep her facial features, face shape, hairstyle, hair colour, skin tone and age exactly as in image 1. She wears a navy blazer over a white shirt and smiles in front of a city office building, natural daylight, shallow depth of field, waist-up framing.

Person plus outfit:

Image 1 is a front-facing photo of a person; image 2 is a beige knit cardigan. Show the person from image 1 wearing the cardigan from image 2, reading by a café window in warm afternoon light. Keep the face and hair identical to image 1, and the cardigan's colour, knit texture and buttons identical to image 2.

Common mistakes

  • Poor reference photos. Profiles, backlight, sunglasses, or a tiny face in a group shot all force the model to guess.
  • Several people in the reference with no indication of who. Write "the man in the white T-shirt in the middle of image 1".
  • A big jump in style. Turning a realistic photo into an exaggerated cartoon loses likeness. Ask it to "stay recognisable".
  • Vague wording about bodies or clothing. It is a common cause of refusals. Name the garment and colour.
  • Using someone else's photo. Use your own photos or ones the person agreed to, and never use them to impersonate someone.

FAQ

How many photos should I upload?

One or two clear front-facing photos of the same person work best; the median on Miku AI was two references. GPT Image models accept up to 16 references and the Nano Banana family 14, but more portraits do not necessarily mean a better likeness.

Why do same-face tasks get refused so often?

The references are almost always real people, so automated moderation is more cautious; about 17% of these tasks were refused. State the purpose ("a professional headshot", "a travel keepsake"), name specific garments, and leave out wording about bodies or revealing clothing. More in why benign requests get refused. Refused tasks are refunded on most models.

The face does not look like the person. What can I do?

Use a sharper front-facing photo and name the features that must match, for example "monolid eyes, round face, a mole at the end of the left eyebrow". If you only need to change the outfit or background of a photo, use "Mask edit" on the reference image to paint just that area, so the face is left untouched; see editing one part and keeping the rest.

How are credits charged?

Per generated image, by model and size tier; the number of references does not change the price. A 1K image is 0.3 credits on the GPT Image models and 0.25 on Nano Banana 2. See pricing.

More posts