Consistent character AI: one face in every post and video

Consistent character AI is any tool or method that keeps one AI-made face the same across many pictures and videos. For you, it fixes one problem in your week: the face. Without one, each carousel cover is a stock photo or a new prompt with a new face, and each Reel means filming or going quiet. With one, the same face turns up on Tuesday's carousel, Thursday's Reel and every post after, and clients know it before they read your name.

That's all it fixes. The character itself isn't the system ^[1]. A reference file solves one problem, a consistent identity ^[1], and the bigger job is still yours every week:

• What to say ^[1]. Take it from the questions clients ask on calls.
• Turning ideas into scripts ^[1], with the video script template for the shape.
• Making usable footage ^[1], the part this page covers.
• Editing, with hooks, captions, B-roll and motion graphics ^[1]. The hook generator gives you openings.
• Posting on time, week after week ^[1].

Nor does it do the trust part. A character gets attention and makes clients feel seen, but it doesn't close ^[3]. It earns reach, and your founder content earns trust ^[3]: you, on camera or in your own words, with the answer to the question the character raised. LinkedIn is on the same side: it welcomes AI posts that reflect a real person's experience, and says generic ones are less likely to be widely shown ^[4]. On weeks you can't film, an AI avatar of you can read scripts you approve ^[3].

Postmosis can make that video from one photo of you: type the script and your photo speaks it, two a day without filming. See AI video

Why is the face the hard part? One good picture shows the model one angle, one lens, one expression and one light, so it guesses the rest and the profile and clothes shift from post to post ^[1]. The fix is to show the model every angle once, and keep that file.

Here's the whole method. Steps one to four are one evening's setup, say a Monday after your last client call, and five to seven take a few minutes each time you post. You can start tonight in a chat model that edits pictures you upload, such as ChatGPT ^[5]: upload your headshot, paste the short sheet prompt from the block at the end, check the panels and save the file. That's the hard part done.

1. Pick one clean picture with the face clearly visible ^[1]: the recent headshot clients know from your profile, or a generated portrait for a made-up client.
2. Turn it into a reference sheet: the same person from ten angles, face and full body, on plain grey, in one outfit ^[1].
3. Check every panel against the original and regenerate any that drift ^[6]. Then fill in the reference card with what a client would notice if it changed, like your glasses or the jacket you wear to meetings.
4. Keep the look of the shot out of the sheet. Light, colour grade and grain go in a separate style reference ^[1].
5. For each new picture, such as Tuesday's carousel cover, attach the sheet, the style reference and a product picture for anything that must stay exact, like your book. Describe the scene and say identity wins when they clash ^[1]. Add the closest angle for a hard shot ^[1].
6. For a Reel, attach the sheet and the closest angle, and say what changes and what stays ^[1]. Test a head turn and a walk once. Keep clips to 4 or 5 seconds ^[6].
7. Save every good shot as a secondary reference for that kind of shot, such as seated at your desk ^[1]. Name it for the post it made and keep it in your own folder beside the sheet, so next week starts from a close match, not a blank prompt ^[1].

Past the setup, each picture is a few minutes plus a careful look against the sheet. Don't skip the look.

There are five ways to get a recurring character. For three posts a week, start with the first, a reference sheet in the chat model you already use, and move on only when it keeps losing the face.

Whichever you pick, keep your own sheet and card. A trained identity stays inside the tool that made it ^[2], and your sheet moves with you to the next one.

Judge a tool on the retries it takes to keep the face, not on its first good picture ^[14]. Give each one the same sheet and the same four shots, a portrait, a full-body pose, a side view and a scene ^[14], taken from next week's real posts. Compare them side by side, count the retries and edits, and keep the tool with the fewest ^[14].

Here's the method on an ordinary week. In another trade, swap the posts and keep the steps.

Here's an ordinary week with the character in it. The pictures come from the sheet in one sitting.

That's three posts a week and one minute of filming. When a week falls apart, the carousel still goes out, since its cover takes minutes. That's the week that used to turn into a quiet month.

Three kinds of recurring character suit a practice of one. An illustrated you fronts carousels on weeks you'd rather not film. A mascot argues with the doubts clients bring to a first call, and the brand mascot playbook builds one. A made-up client acts out the problem you solve, so the client watching sees themselves. Start with one, and add a second only for a clearly different client group ^[6].

Keep the formats apart. A character lives your client's problem, a mascot is your brand beating their objections ^[6], an avatar is you reading your scripts, and UGC and ads sell a product directly ^[3].

Some posts only work with you in them. Film or write these yourself:

• The first time a prospect meets you. Pin a video you filmed yourself to your profile, so the first face they see is the one they'll meet on the call.
• A client's painful moment, such as a divorce or a death in the family. Played by a cartoon, it looks like mockery.
• Posts built on a result, like a won case or a sale price, because the character doesn't promise results ^[3] and those posts need you to stand behind them.
• Weeks with no time to check each picture against the sheet. A face that changes between videos never becomes a brand ^[6]. Post a plain text carousel instead.

For most coaches, consultants, agents, freelancers and lawyers, the most useful recurring character is you, drawn. It fronts carousels without a camera and still looks like the person a client meets on a call.

1. Pick a recent, well-lit headshot. Only use photos of yourself or of someone who's agreed ^[2].
2. Describe one style closely, such as line weight, colours and shading, since specific details give more consistent results ^[7]. A new style later means a new sheet.
3. Run the illustrated-you prompt from the block at the end. Regenerate until it looks like you at a glance, then stop.
4. Build the ten-angle sheet from that portrait ^[1], and write on the card what makes it you: hair, glasses, the jacket you wear to meetings.

Judge the portrait against a recent photo, not how you'd like to look, since a flattering drawing backfires once a client turns their camera on. New glasses or a new haircut mean a new portrait and sheet.

Put the drawn you on the slides: the Postmosis carousel maker turns a post, pasted text or a topic into 2 to 10 slides in your brand colour, and downloads them as pictures or one PDF for LinkedIn. See the carousel maker

For a made-up client or a mascot, decide who the character is before you make an image. An AI character is a made-up person who turns up in every video with the same look and the same running joke ^[3]. For a practice of one, the useful version lives your client's problem: people watch because it's funny, and they keep watching because it's them.

Four things make one easy to recognise ^[3]: a look you'd know in half a second, the same hair, outfit and face every time, one behaviour, ideally a format that already works, and one thing that changes, usually the place ^[3]. The brief is six decisions, in order, and the template at the end has a line for each:

1. The client's fear or frustration, in their words from reviews, calls and your inbox. Pick one. A character with three problems has none ^[3].
2. The character, as a casting note: age and vibe, one hair detail, the same outfit every time, an optional prop and a default expression. A silhouette should be enough to recognise them, and they're never based on a real person, a celebrity or an existing character ^[3].
3. The one behaviour, repeated in every video, that shows the fear without a word. If it needs a caption to make sense, simplify it ^[3].
4. Five situations to start with. The behaviour stays the same in all of them, and only the place or the moment changes ^[3].
5. The link. One link, one action, matched to the fear: a nervous character points to something gentle, such as a page for nervous first-timers instead of a price list ^[3].
6. The handover: a character earns attention but not trust ^[3], so decide who the real person is, where they appear and what they say first.

The same six answers work for any practice. A dentist's version is a nervous patient who freezes at the door, with a booking page called "Nervous patients start here" ^[3]. The trade table above has two more starting points: an agent's buyer who keeps losing and a fitness coach's Monday quitter ^[3].

The default AI face, with perfect skin and hair, is forgettable ^[6]. An odd detail stops a scroll, such as a strange haircut or oversized glasses, but it still has to suit the client you're after ^[6]. Decide who's watching, what they'd follow (a mentor, a funny friend, a sceptic) and what the character always does. Then run the five-concepts prompt at the end with your brief and a few mood pictures.

Generate the portrait in an image model and keep the one you'd still recognise in a crowd ^[6].

Make the sheet once, in one sitting. It's one image of the character from every angle, and every later carousel and Reel points back to it ^[1]. You want an accurate record of one face, so don't let the model make it prettier. Keep the rear view in, because a video model needs it to turn the character ^[1]. Both sheet prompts are in the block.

Check every panel against the original before you move on ^[6], hardest at the profiles and the back, since your one photo never showed the model those angles ^[1]. Then fill in the card, the sheet's written half. You'll paste lines from it into every prompt. It also says what may change between Tuesday's carousel and Thursday's Reel: the place and the props, never the face or the glasses.

Every picture you post asks the model four questions. Keep the answers in separate files and consistency gets much easier ^[1].

Keep the sheet neutral. Cinematic light, a heavy colour grade, grain or a location baked into it makes the identity less reusable ^[1]. For a look, such as the warm window light you'd like on every cover, run the style prompt on a photo that has it. Take the photography and nothing about the person, so Tuesday's carousel and Thursday's Reel look like one series.

For Reels, the sheet is an identity library, so you don't upload every panel ^[1]. Attach the full sheet, plus the closest angle for a hard shot, a location picture if the framing matters, and a product reference if something must stay exact ^[1]. Before any real episode, run the head-turn and walk tests from the block in an image-to-video model such as Seedance.

Once the face holds, build episodes from short clips. Write each scene yourself from a still, or use motion transfer so your character performs a format that already works ^[6]. Borrow the movement and the structure, and make your own base clip. Never repost someone else's video with a new face ^[6]. TikTok also keeps reused material with nothing new out of its For You feed ^[16].

Cut the clips together, then add your voice, captions and music. Plan retries on the first episodes, and keep a phone video in reserve for the week a clip won't hold.

Some shots still come back wrong, and a client who sees your posts every week will notice. Here's the cause and the fix.

Know what it can't promise. Even a trained identity gives you clearly the same person, not a pixel-identical face ^[2], so check every result. If a shot drifts after a few tries, change the shot, not the prompt: a closer angle or a simpler pose. A whole batch drifting means rebuilding the sheet from a cleaner picture.

Decide what the attention is for before the first video ^[6]. Usually it's your own service, with the character as its face, or a free guide that leads to a consult.

1. One call to action per video, such as commenting a word to get your guide ^[6].
2. Reply to everyone who comments that word with the guide, and make the guide lead to a booking page or a consult.
3. Put the same link in your bio for everyone who doesn't comment ^[6]. A link in bio page that counts clicks per link shows whether it gets used.

Then the handover: the real person, usually you, answers the fear the character acts out, behind the link or in a pinned video ^[3].

Treat the first ten videos as a test of the character rather than of sales ^[3]. The signal is people commenting that the character is them ^[3]. Until then, change the situations and keep the face. After that, track keyword comments per 1,000 views, and if that's low, change the call to action before the character ^[6].

Don't guess what the character does next. An outlier is a video that gets far more views than its account usually does ^[6], which tells you the format works right now.

1. List 20 to 40 accounts in your field, and put some much smaller than you on it ^[6]. A family lawyer's list might be other family lawyers and estate planners.
2. For each, compare recent videos with the account's usual views. Anything at 5x or more is an outlier ^[6].
3. Write down the format: the hook, the length, the setting and what's on screen ^[6].
4. Put your character into that format, with your own angle.

Four accounts to start your list, as format references whose results haven't been verified ^[18]. A dentist, Smilestudio (@smilestudio__ss), answers a patient question with cartoon teeth fighting for space. Brand designer Denny Kurien (@dkurien) records his voice on one idea and gives each sentence its own picture. Wei Chen (@wealthbywei) keeps one role, one format and one offer, but appears to be AI with no label, so label yours ^[18]. Bailey and the Bean (@thebaileytales) run episodes in seasons, with one link in bio doing the selling ^[18].

A made-up client or a mascot needs more setup than an illustrated you, so spread it over week one, with three videos a week on days you can keep:

Three a week that you keep beats a burst of daily posts and a quiet month.

Put the plan on the Postmosis content calendar and get a phone alert when each post is due, or schedule it and let Postmosis post for you once your accounts are connected. See the content calendar

• Say it's AI, in the bio and with each platform's label ^[19][16]. AI images don't need Instagram's label ^[19], but 'Illustrated with AI' in the caption costs nothing, and LinkedIn asks you to say so when heavy AI use isn't obvious ^[4].
• No advice, promises or testimonials from the character ^[3]. A made-up mentor can't claim a track record or clients ^[6], and the FTC's fake-review rule covers people who don't exist ^[15].
• Laugh with your client, since the joke is the situation everyone recognises and never the person ^[3].
• No real likenesses: not a client, a competitor, a celebrity or a well-known cartoon ^[3].
• Check your profession's advertising rules first. Dental, medical, legal and financial work often has its own ^[3].
• If anyone pays you to feature them, use #ad or the paid-partnership label ^[6].
• Keep a real person easy to reach from every post.

Everything you paste, in the order you'll use it.

Monday evening, your first sheet. Upload one clear picture, then paste this.

Use the uploaded image as the only source for this person's identity. Make one character reference sheet with ten panels on a plain light-grey background, in the same outfit and the same soft, even light.

Top row, face: close-up front, chest-up front, chest-up 45 degrees left, chest-up 45 degrees right, chest-up left profile, chest-up right profile.
Bottom row, full body: front, three-quarter, side, back.

Don't redesign or beautify them. Rebuild one person and rotate them, so it looks like one real person photographed from ten angles in one session.

The quick prompt for each post's picture.

Use the attached reference sheet as the only source for who this person is: same face, hair, body and outfit. Use the attached style reference only for lighting, colour, grain and lens. Scene: [WHAT HAPPENS IN THIS POST]. If they conflict, the identity wins.

For the drawn you on your covers. Attach a recent headshot.

Turn the attached photo of me into an illustrated portrait for my social posts. Keep me recognisable: face shape, eyes, nose, hairline, skin tone, hair and glasses. Style: [for example, flat illustration, bold black outlines, simple cel shading, navy, teal and warm grey]. Chest-up, facing the camera, friendly neutral expression, plain light background, no text. Don't make me younger, thinner or more polished than the photo.

For a made-up client, or a mascot for your practice.

CHARACTER BRIEF

1. The client's fear or frustration (their words, one only):
"..."

2. The character
Age and vibe (one line):
Hair (one unmistakable feature):
Outfit (the same in every video):
Prop (optional):
Default expression:
Not based on: any real person, celebrity or existing character

3. The one behaviour (shows the fear without a word):

4. Five situations (only the place or moment changes):
a.
b.
c.
d.
e.

5. The link (one action, matched to the fear):

6. The handover
Real person:
Where they appear (page behind the link, pinned video, comment reply):
First thing they say:

7. Labels
Bio line saying the account is AI:
Platform AI label switched on: yes / no

Paste it with your brief and a few mood pictures.

I'm building a recurring AI character for my business. I'm a [what you do] and the character is for [one client, described in a sentence]. It should lead them to [your offer or free guide].

Here's my character brief:
[PASTE BRIEF]

Use the reference images I've attached for mood only. Give me five character concepts. For each one: a name, the look (face, hair, outfit, one unusual detail), the personality in one line, the thing they do in every video, and why my client would keep watching. No real person's likeness.

Then write an image prompt for the strongest concept: front-facing portrait, plain background, soft even light.

The full sheet, for when the short one drifts.

Create a professional multi-angle CHARACTER REFERENCE SHEET using the uploaded image as the identity reference.

The person in the uploaded image is the only source for the character's identity. Don't redesign, beautify or reinterpret them. Reconstruct the same person as one consistent three-dimensional identity and show them from standard camera angles.

Identity consistency is the top priority.

Make ONE reference sheet with ten panels:

TOP ROW: FACE
1. Close-up, straight-on face
2. Chest-up, straight-on
3. Chest-up, about 45 degrees left
4. Chest-up, about 45 degrees right
5. Chest-up, exact 90-degree left profile
6. Chest-up, exact 90-degree right profile

BOTTOM ROW: BODY
7. Full body, straight-on front
8. Full body, about 45-degree three-quarter view
9. Full body, exact 90-degree side
10. Full body, exact rear view

Plain light-grey studio background, soft even frontal light, minimal lens distortion, and the exact same outfit in every panel.

Don't generate the person separately for each panel. Rebuild one person and rotate them. Keep the same facial geometry, nose shape, jaw, eye placement, hairline, skin tone, body proportions, apparent age and outfit in every view.

It should look like one real person photographed from ten angles in one studio session.

Fill it in once, then paste lines from it into every prompt.

CHARACTER REFERENCE CARD

Name:
Reference sheet file:
Original portrait file:

Identity (never changes)
Face: [shape, eyes, nose, mouth, distinctive marks]
Hair: [colour, cut, the one detail]
Build:
Outfit: [the signature outfit, item by item]
Glasses or prop:
Default expression:

Voice (for video)
Voice name or clone:
Pace and tone:
Running joke or catchphrase:

Style references (kept separate)
Style file 1: [what it's for, such as daylight interiors]
Style file 2:

Best shots so far (secondary references)
Carousel cover, pointing at the question:
Seated at my desk, talking to camera:
Side profile:
Full body, walking:
Outdoors in daylight:

What may change: [place, time of day, props in the scene]
What must never change: [face, hair, outfit, glasses]

Run it on a photo whose look you want on every post.

Analyse this image as a reusable photographic style reference.

Return one JSON object that describes only the visual treatment, so I can apply the look to future images of different people and products. Keep the style separate from the subject's identity.

Cover:
- light: direction, softness, contrast, light sources, exposure, shadows and black levels
- colour: white balance, the main colours with approximate hex values, saturation, and the grade in highlights, midtones and shadows
- skin rendering, without describing the person
- camera: lens, likely focal length, depth of field, camera height, framing and composition
- texture: grain, sharpness, background treatment, and any bloom, flare, diffusion or motion blur

Don't describe or copy the person's face, body, clothing, ethnicity, age or identity.

End with a short "style_prompt" field I can paste into an image prompt.

The full prompt for each new picture, such as Tuesday's cover.

Use the attached character reference sheet as the only source for identity.

Keep the exact same person: facial geometry, eye shape and spacing, nose, lips, jaw and chin, skin tone and texture, hairline, hair texture, body proportions and apparent age.

Use the attached style reference (image or JSON) only for the photographic treatment: lighting, colour grading, contrast, grain, lens, depth of field, composition and general look. Don't take the person, their features, body or clothes from the style reference.

Scene:
[WHAT HAPPENS IN THIS POST, such as me at a whiteboard for this week's carousel cover]

If a product picture is attached, keep its exact shape, proportions, packaging, logo position and colours.

Priority when these conflict:
1. character identity
2. product accuracy
3. scene
4. photographic style

It should look like a real photograph of the same person in a new situation, not someone similar.

Both tests, once, before your first Reel.

Use the attached character reference as the exact identity.

Start with [CHARACTER] in a clean left profile beside a large office window. They look outside for a moment, then turn their head and upper body naturally toward the camera until they reach a three-quarter view.

Slow, subtle camera push-in. Soft, realistic window light. Natural blinking, subtle breathing, realistic skin texture and gentle hair movement.

The nose, jawline, eyes, cheekbones, hairline, skin tone and facial proportions stay the same as the head turns. Don't morph, beautify or redesign the face during the turn.

Use the attached character reference as the exact identity.

Full-body shot of [CHARACTER] walking slowly toward the camera through a bright, simple office. Keep the same face, hairstyle, body proportions and outfit as the reference.

They walk naturally, glance away, then look back at the camera with a small smile.

The camera tracks slowly backward at eye level. Natural daylight, realistic body mechanics, subtle hair movement, no exaggerated gestures. Keep the identity consistent for the whole shot.

One clip of Thursday's Reel, from a still of the character.

[CHARACTER NAME] from the reference image, [SETTING, such as your office, a client's kitchen or an open house], [ONE ACTION] while talking to the camera. Handheld phone framing, natural light, 5 seconds. Keep the face, hair and outfit exactly as in the reference.

Tool settings worth knowing:

• Midjourney's Edit model: up to 4 reference images ^[8].
• Higgsfield Soul ID: 20 or more recent photos of one person, up to 80, with at least one full-height shot ^[2]. It carries into Seedance video as an Element, and it can't be exported ^[2].

Run each new batch down this before it goes on a post.

[ ] One neutral master sheet, checked panel by panel against the original
[ ] Reference card filled in: what never changes, what may
[ ] Identity and style kept in separate references
[ ] Style taken from a photo with the look you want, never the person in it
[ ] Closest angle attached for hard shots, such as a profile or a walk
[ ] Video prompts say what changes and what stays
[ ] Clips kept to 4 or 5 seconds, cut together in the edit
[ ] Best shots saved back as secondary references
[ ] New shots start from the library, not from a blank prompt