The five-angle reference system I use to keep six characters recognisable across eleven episodes of a series — and one AI singer recognisable to 92,000 followers.
No signup. Nothing gated. Read it, build the sheet, use it today.
START HERE
Most people respond to a character that keeps changing by switching tools. Then it happens again in the new tool. The model was never guessing — it was answering exactly what you handed it, and what you handed it was incomplete.
A good reference is a brand decision that can be tested, improved and made accurate.
A reference is not a mood board. It is an instruction about what must stay the same. Everything below is how to write that instruction so the model cannot misread it.
Before any image exists, write the short list of things that are never allowed to change. Not a description of a person — a list of constraints.
Face shape. Approximate age. Build. Hair length and colour. One signature detail that is hers alone — a scar, a chain, a jacket, a way she holds herself.
Five to seven items. If your list runs to twenty, you have written a character description, not a lock. The model will average a long list and give you a stranger.
WHY THIS COMES FIRST
A reference image can only protect the details you already decided matter. Undecided details drift by definition — the model fills them in freshly every single time.
One image becomes the source of truth. Everything else in your project is generated from it or checked against it. I build this in ChatGPT Images or Nano Banana.
Neutral lighting, neutral expression, plain background, front on. Not the most beautiful image you can make — the most legible one. This frame exists to be read by a machine, not admired by a person.
Generate it more than once. Compare candidates against your list from step one, not against your taste. The one that holds every constraint wins, even when a prettier option exists.
One image is not a lock. Every shot you will ever generate pulls from a different angle, and a model handed only a front view has no information about the rest — so it invents it. That invention is the drift you are seeing.
Five views cover everything you will ever ask for. The prompt below produces all five in one pass, from your Master Reference.
THE PROMPT
Works in ChatGPT Images or Nano Banana. Paste your list from step one into the block marked MUST NOT CHANGE. One pass, five views, one image.
Using the attached image as the single source of truth for this character, generate a five-view character reference sheet as ONE image. VIEWS, left to right: 1. Front - straight on, eyes to camera 2. Three-quarter - head turned 45 degrees 3. Profile - full side, 90 degrees 4. Half body - waist up, front on 5. Full body - head to feet, standing, arms relaxed MUST NOT CHANGE between views: [paste your list from step one here] IDENTICAL IN ALL FIVE: - Lighting: soft, even, frontal. No shadow drama. - Expression: neutral. Mouth closed, eyes open. - Clothing: exactly the same garment. - Background: one flat colour, no environment. - No filter, no grade, no stylisation. OUTPUT: One image. Five views in a row, equal height, evenly spaced, same scale relative to each other.
This sheet is a measuring tool, not a portfolio piece. Every constraint above exists to remove a variable — anything that changes between the five views is a variable the model is free to keep changing later.
Regenerate view [NUMBER] only. Match the attached sheet exactly - same person, same lighting, same clothing, same scale. MUST NOT CHANGE: [paste the same list] Change nothing else. Output only the one view.
Look back. If any one of them reads as a different person, run the repair prompt before you produce a single scene. A grid you have not audited is not a lock — it is five images that happen to look related.
WHERE THIS STOPS
Everything above solves a technical problem completely. Do it properly and your character will hold across a hundred shots, eleven episodes, a year of posting.
A locked character is still a stranger. Consistency makes her recognisable. It does not make anyone care.
Read the prompt again and notice what it never asked you. It never asked who this character is. It never asked why a stranger should follow her, what world she lives in, or what happens in the next post that makes anyone wait for it. It could not ask — those are not things a model can hold for you.
This is the wall people hit about three weeks in. The face finally holds. The views do not move. And the honest diagnosis is that the face was never the reason nobody was watching.
If the only reason to follow your character is that it was made with AI, you built a gimmick. If people want to know what she does tomorrow, you built a brand. That gap — identity, a world, a reason to return, a series people wait for — is what the full method covers.
Five guides. How to find what is worth producing, how to direct a moment so it lands, how a character becomes an identity people follow, how to turn your world into a series, and how a commercial brief becomes an idea worth making.
You have the lock. Next is the reason.
See the full method → Built from Monica — 92,000 followers — and Only in Israel, eleven episodes and over a million views.