AI Headshot Prompts: What Actually Works, and What Doesn't

Most articles on this topic hand you thirty prompts that differ only in jacket color. This one is shorter, because the useful part isn't the prompt list — it's understanding which parts of a prompt the model vendors actually document as working, and which parts are copied folklore that does nothing.

The short version: describe the shot the way a photographer would brief one, edit your own photo instead of generating a face from scratch, and repeat what you want preserved on every single iteration. Everything below is either quoted from vendor documentation or flagged as unverified.

Start from the right job: editing, not generating

Nearly every prompt list in this category is written for text-to-image generation — describe a person, get a person. But almost everyone searching for this already has a selfie and wants that face, improved.

Those are different tasks with different prompts, and the vendors document the difference clearly.

Google's image generation prompting guide gives a template specifically for editing an existing image:

"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."

Google Cloud's Nano Banana prompting guide says the same thing as a mindset shift: "Editing requires a different mindset than generating. You already have a base image; your prompt needs to focus on what is changing and what is staying the same."

OpenAI's image generation prompting guide adds the operational detail that matters most, and that no competing article mentions:

"For edits, use 'change only X' + 'keep everything else the same,' and repeat the preserve list on each iteration to reduce drift."

Repeat the preserve list every time. That's the single highest-value instruction on this page. Likeness drift across iterations is the reason people end up with an attractive stranger after six rounds of refinement.

The formula, from the vendor that published it

Every "five-element framework" in this category is a reinvention of a template Google published. Here it is in full, from Google's own guide:

"A photorealistic [shot type] of [subject], [action or expression], set in [environment]. The scene is illuminated by [lighting description], creating a [mood] atmosphere. Captured with a [camera/lens details], emphasizing [key textures and details]. The image should be in a [aspect ratio] format."

Google's stated principle behind it is worth internalizing, because it contradicts how most people write prompts:

"Describe the scene, don't just list keywords. The model's core strength is its deep language understanding. A narrative, descriptive paragraph will almost always produce a better, more coherent image than a simple list of disconnected words."

And on why photographic vocabulary belongs there at all: "For realistic images, think like a photographer. Mentioning camera angles, lens types, lighting, and fine details will guide the model toward a photorealistic result."

OpenAI recommends a consistent ordering — "background/scene → subject → key details → constraints" — and suggests including the intended use, which "set[s] the 'mode' and level of polish."

What's documented, what's caveated, and what's folklore

This is where most prompt lists mislead people. Three tiers:

Documented as working

Documented with a caveat — the 85mm question

Everyone writes "85mm, f/1.8" and nobody explains it. OpenAI does:

"detailed camera specs may be interpreted loosely, so use them mainly for high-level look and composition rather than exact physical simulation."

So "85mm portrait lens" reliably pulls the model toward the look of portrait photography — compressed features, separated background, soft falloff — because that's what images labeled that way look like in training data. It is not simulating a lens. The inference — mine, not the vendor's — is that the specs don't stack: writing "85mm, f/1.8, 1/200s, ISO 100" gets you roughly what "85mm portrait lens" already got you. If the specs are being read loosely, piling on more of them can't be read any more tightly.

Both things are true at once — Google validates the vocabulary, OpenAI tells you not to over-read it.

Folklore

• "8K", "ultra-detailed", "high resolution." Actively counterproductive as a substitute for real description. OpenAI's gpt-image-1.5 prompting guide puts it directly: "For photorealism, camera/composition terms (lens, aperture feel, lighting) often steer realism more reliably than generic '8K/ultra-detailed.'" Resolution is a parameter you set, not a word you type.
• The "no smoothing, no airbrushing, no plastic finish" stack. The intent is right and OpenAI endorses it — but the phrasing fights documented guidance. Google: "Use 'semantic negative prompts': Instead of saying 'no cars,' describe the desired scene positively." Google Cloud repeats it as a top-four best practice. Write "natural skin with visible pores, fine lines, and true-to-life color" instead.
• "Rembrandt lighting," "three-quarter crop," "catchlight in both eyes." These appear in every list in this category. None of them appears in any vendor documentation. They may work — they're plain descriptive language, which is documented — but not because they're recognized controls. Use them without believing they're magic words.

The preserve block

Append this to any headshot edit, and repeat it on every follow-up. It's assembled from the documented patterns above rather than invented:

Keep everything else exactly the same. Preserve my exact facial structure,
eye shape and color, nose, jawline, hairline and hairstyle, skin tone, and
any distinguishing features such as freckles, moles, or scars. Keep natural
skin texture with visible pores and fine lines. Do not slim, smooth, or
idealize my face. Do not change my apparent age.

The face-shape and age lines matter more than they look. Drift across iterations is documented by OpenAI, which is why it recommends repeating the preserve list; the direction of that drift — toward a smoother, younger, more conventionally attractive average — is my own observation rather than anything a vendor states. Either way it happens gradually enough that you won't notice until you compare against the original, so keep the original open.

If likeness starts slipping anyway, Google's documented fix is to stop patching: "If you notice a character's features begin to drift after many iterative edits, you can restart a new conversation with a detailed description to retain consistency."

Eight prompts that are actually different from each other

Prompt 1 is complete. Prompts 2 through 8 each start with ..., which means: keep prompt 1's opening clause ("Using the provided photo, restyle it as a photorealistic professional studio headshot") and swap in what follows. Add the preserve block to all of them.

1. Standard corporate — the safe default.

Using the provided photo, restyle it as a photorealistic professional studio headshot, framed chest-up against a smooth medium-grey background. A large soft key light slightly above eye level with gentle fill creates soft catchlights in the eyes. Captured with an 85mm portrait lens at f/2.8 with a softly blurred background. Relaxed, confident expression. Business-casual navy jacket over a plain shirt.

2. Bright and approachable — for client-facing and education roles.

...set against a softly blurred bright interior with warm daylight from a window at 45 degrees to the left. Warm, natural expression with a real smile. Mid-tone knit or soft blazer.

3. Executive — authority without coldness.

...framed slightly tighter, chest-up, against a dark neutral gradient. A single key light with controlled falloff and deep, soft shadow on the shadow side. Composed, direct expression, no smile. Well-fitted charcoal suit with a plain shirt.

4. Environmental — for a bio page rather than a profile square.

...standing in a real workplace with the background thrown well out of focus, lit by available daylight plus soft fill. Captured with a 135mm lens at f/2, medium shot from the waist up, subject positioned off-center with space to look into.

5. Outdoor natural light.

...outdoors in open shade with soft, even light and a distant blurred treeline. No harsh sun on the face, no dappled shadows. Captured with an 85mm lens at f/2.

6. Creative or editorial — for design, media, and founders.

...against a textured neutral wall with directional side light and defined shadow. Higher contrast, cooler color grade. Head slightly turned, gaze direct.

7. Black and white.

...converted to a high-quality black and white portrait with full tonal range, deep blacks, and clean highlights. Directional soft key light. No color cast, no heavy grain.

8. Cleanup only — the least risky option, and often the best.

Using the provided photo, do not regenerate the image. Apply post-processing only: correct the white balance, even out the exposure, gently lift shadows, reduce noise, and sharpen the eyes. Neutralize the background to a clean even tone. Maintain the exact original facial structure, expression, clothing, and framing.

That last one deserves emphasis. If your source photo is decent, a cleanup prompt gets you a better and more honest result than any generation prompt, and there's nothing for likeness to drift away from. Start there before you escalate.

Fixing the four things that go wrong

It doesn't look like you. Your reference photo is doing the heavy lifting. Feed it a well-lit, front-facing image at a decent resolution rather than a dim group-photo crop. Then strengthen the preserve block and change one variable at a time — OpenAI's advice is to "start with a clean base prompt and refine with small, single-change follow-ups."

The skin looks like plastic. Add texture positively rather than negatively, and check the model's quality setting. OpenAI specifically recommends comparing higher quality settings for "close-up portraits, identity-sensitive edits."

The crop is wrong for LinkedIn. Profile photos display in a circle, so anything near the frame edge gets cut. Specify the framing in the prompt and set a square aspect ratio, then check the result inside a circular mask before uploading.

The model refuses. This is normal and expected. Midjourney warns users directly: "You will likely encounter friction with our moderation — seemingly innocent prompts may be blocked by our filters. Blocked jobs don't cost you any credits." Rephrase descriptively rather than trying to route around the filter.

Consent, provenance, and the rules nobody mentions

Not one article ranking for this keyword mentions any of this, and all of it comes from the vendors' own terms.

You need consent for someone else's face. OpenAI's Service Terms state: "You may not use Visual Capabilities to reproduce the likeness of any person without express consent and all necessary rights." Their usage policies prohibit "use of someone's likeness, including their photorealistic image or voice, without their consent in ways that could confuse authenticity." Midjourney's Omni Reference documentation puts it plainly: "You must have the necessary rights to use the images you upload."

Your own selfie is fine. Generating headshots for your whole team from photos you pulled off the company directory is not — and that's a terms-level obligation, not etiquette.

Don't claim it's a camera photo. Google's Generative AI Prohibited Use Policy prohibits "Misrepresenting the provenance of generated content by claiming it was created solely by a human, in order to deceive," alongside "Impersonating an individual (living or dead) without explicit disclosure, in order to deceive."

It's labeled whether you mention it or not. Google Cloud states that its image outputs carry "C2PA Content Credentials and a SynthID watermark." The provenance travels with the file.

Check the platform's own rule before you upload. LinkedIn maintains its own profile photo guidelines, and that page — not a blog post, including this one — is the authority on what it will and won't accept. Read it before you swap your photo. Whatever it says, a result that still looks like you is the one least likely to cause a problem, which is another argument for the preserve block.

Midjourney is the wrong tool for this

Worth saying plainly, because it appears in every prompt list in this category. Midjourney's own character reference documentation states: "For best results, start with an image of a single character created by Midjourney. Images of real people typically won't look exactly like them." And elsewhere: "Midjourney uses Image Prompts and references as inspiration to guide new creations, not to copy them exactly."

It's an excellent tool built for a different job. If you want a headshot of you, use something built around identity preservation.

(Note also that most prompt lists still show Midjourney's --cref parameter. Midjourney's own docs now direct V7 and later users to Omni Reference instead, so that advice is out of date wherever you find it.)

How this works in 43frames

Our headshot templates are prompts — literally. Picking one loads its full prompt text into an editable field, so you can read exactly what's being asked for, change any part of it, and save your edited version as your own template.

Here's the Professional Headshot template's prompt, in full, so you can see the pattern applied:

A professional studio headshot photograph of a person facing the camera with
a relaxed, confident expression, framed chest-up against a smooth medium-grey
background. A large soft key light slightly above eye level with gentle fill
creates soft catchlights in the eyes. Shot on an 85mm portrait lens at f/2.8
with shallow depth of field. Natural skin with visible pores and true-to-life
color, no heavy retouching. If a photo of a person is provided, keep their
exact face, hairstyle, and skin tone. Do not add any text, watermarks, or logos.

Every documented element is there: shot type, subject and expression, framing, background, lighting description, camera and lens, texture requirement, and a preserve instruction. Copy it into any tool you like — the structure is what matters.

There's a separate negative-prompt field alongside it, and given Google's semantic-negative guidance above, that's the right place for the handful of things genuinely worth excluding — text, watermarks, extra fingers — rather than stuffing "no plastic skin" into the main prompt where it competes with your positive description. There are role-specific variants in the template library too, including teacher headshots, lawyer headshots, and executive headshots, each with the wardrobe, background, and expression slots filled differently.

If you're still deciding whether to generate at all, AI headshots vs a photographer covers the trade-off, and our comparison of AI headshot generators covers the tools. And if you'd rather just take a good photo — which, with a window and a tripod, is often the better answer — professional headshots at home is the guide for that.

FAQ

Does "85mm lens" actually do anything? Yes, as a style token. Google documents lens terms as composition controls; OpenAI adds that camera specs "may be interpreted loosely." It steers the look, it doesn't simulate a lens.

Should I write "no plastic skin, no airbrushing"? Probably not in that form. Google's guidance is to describe what you want, not what you don't. "Natural skin with visible pores and fine lines" follows the documented pattern.

Does adding "8K" improve resolution? No. Resolution is an output setting, not a prompt word.

Can I generate a headshot of a colleague? Not without their consent — OpenAI's Service Terms require express consent and all necessary rights for anyone's likeness.

Is Midjourney good for this? Not for a headshot that has to look like you. Its own docs say "Images of real people typically won't look exactly like them."