Most people don’t have a prompting problem. They have a reference problem. You write the same description twice, generate twice, and get two people who look like cousins. Face drifts. The jawline softens. The wardrobe changes on its own. Here’s the workflow we use at Zip AI to make the same character show up in every single shot — not a family resemblance, the same person.
This is the exact process behind our own on-camera roster, and it takes about twenty minutes to set up per character. After that, consistency is free.
Why prompt-only approaches always drift
A text prompt is a description, not an identity. “Adult Black man, close-cropped hair, short beard, charcoal suit” narrows the field to maybe ten thousand plausible faces — and the model picks a different one every run. Add more adjectives and you narrow it further, but you never close the gap, because you’re describing a type, not a person.
Consistency comes from giving the model an actual visual anchor to match against, then writing rules that stop it from averaging that anchor with anything else. Two parts. Skip either one and you get drift.
First decision: Element or Soul?
Higgsfield gives you two ways to lock a character, and picking the wrong one costs you an afternoon.
Use an Element when you need a reusable reference fast, when more than one character appears in the same shot, or when your subject isn’t a person at all. It’s built from a single image, it’s ready immediately, and you can stack several references in one generation.
Use Soul when you’re building a true digital twin of one specific person and you have 5–20 photos of them. It takes about ten minutes to train and handles one person per output.
For most brand work — a recurring spokesperson, a team of characters, anything with two people in frame — Elements is the right call. Everything below assumes Elements.
What you need before you start
Four reference images per character, and a written identity lock. That’s it. Don’t start generating until you have both.
Step 1 — Build the four canonical references
Generate or shoot four images of your character and keep them somewhere permanent. Each one does a specific job:
- Canonical front — the authority image for face and overall presence. This is the one you’ll compare everything against.
- Three-quarter — controls how the face holds up when the head turns. Most drift shows up here first.
- Close portrait — controls fine facial detail: eye spacing, hairline, beard shape, skin texture.
- Full body — controls proportions and height only. Note this one clearly, because if it’s shot in casual clothes it will try to override your real wardrobe.
The fastest way to produce a clean set is a split-screen character sheet: a full-body standing shot on the left, a tight chest-up portrait on the right, on a pure white seamless background with soft diffused lighting. Ask for visible skin texture and natural asymmetry explicitly — pores, fine lines, slightly uneven makeup, matte complexion, no beauty filter, no digital smoothing. Without that, you get the airbrushed plastic look that reads as AI from across the room.
One more thing that matters: for adult characters, say so structurally. “Defined jawline and cheekbones, mature adult bone structure, not a youthful rounded babyface.” Models bias young unless you stop them.
Step 2 — Create the Element and record its ID
Upload your canonical front image and create an Element from it. Give it a real name you’ll recognize a year from now — we use FirstName-LastName-Year — and write down both the name and the ID it returns.
Keep this in one file per brand. Ours is a continuity registry: character name, role, Element name, Element ID, and the file paths to all four reference images. When a new person joins the production, that file is the handoff. No tribal knowledge, no guessing which face was the real one.
Step 3 — Write the identity lock
This is the part almost everyone skips, and it’s the part that actually works. Write a short, blunt list of traits that are never allowed to change, then paste it into every prompt for that character. Ours reads like this:
Preserve face geometry, complexion, apparent age, hairline, beard shape, and body proportions across every shot. Never average this character with another person or another Element. Do not use any other roster identity as a face or body reference.
That last line matters more than it looks. Once you have more than one character on the books, models will happily blend them — and a face that’s 80% the right person is worse than one that’s obviously wrong, because you won’t catch it until the client does.
Step 4 — Lock the wardrobe separately
Faces are sticky. Clothes are not. Treat wardrobe as its own lock and describe it head to toe with materials, cut, and color named: top, layers, bottom, belt, shoes, jewelry, bag. If there’s no bag, say “no bag.” A gap in the list is an invitation for the model to invent something.
Then add the exclusions. “Do not add a tie. Do not change the neckline. Do not swap the jacket between shots.” Negative instructions are doing real work here — the wardrobe drift you don’t forbid is the wardrobe drift you’ll get.
Step 5 — Generate and verify
Generate your shot, then put it side by side with the canonical front reference and check four things in this order: face geometry, hairline and beard, wardrobe, body proportions. Drift almost always appears in that sequence.
What drift looks like in practice: the jaw gets softer and rounder, the character reads five years younger, the beard changes length, or the suit picks up a texture that wasn’t there. Any of those means the reference isn’t weighted strongly enough — restate the identity lock and regenerate rather than trying to fix it downstream.
Common mistakes
- Using the casual full-body shot as the main reference. It’ll pull the wardrobe with it. Full body is for proportions only — label it that way in your notes.
- Letting “realistic” mean “flawless.” Retouched skin is the single biggest tell. Ask for pores, asymmetry, and a matte finish every time.
- Re-describing the character from memory on shot 12. Copy and paste the identity lock. Every detail you retype is a detail you might change by accident.
- Mixing references across characters. One Element per person, always. Never stack two character references hoping for a blend.
- Not writing the ID down. Six weeks later you’ll have four similar Elements and no idea which one shipped.
Put it to work
Set this up once per character and the payoff compounds — every ad, every explainer, every social cut uses the same person, and your brand stops looking like it hired a new spokesperson every month.
If you want the templates we use for this — the continuity registry, the identity lock, and the wardrobe lock — they’re in the Zip AI Operator Kit. And if you’d rather we just build the roster for you, start with a free Blueprint and we’ll map it out.