Building an AI Persona for Instagram That Does Not Change Face Every Post

The first ten posts are easy. You write a prompt, you get a striking face, you post it. The eleventh is where it falls apart: the jawline moved, the eyes are a different colour, and the followers who came for a person are looking at her cousin.

People usually respond by writing longer prompts. More adjectives, a fixed hair colour, a named eye shade, a numbered age. It does not work, and it cannot: a text prompt describes a type of person, never a specific one. Consistency is not a writing problem. It is a reference problem.

Step one: cast the face once, then stop casting

Spend one session generating faces rather than posts. No background, no outfit, no story - just head and shoulders, a neutral setting, several angles. Keep going until one of them is the person. Then stop generating faces entirely.

Save three to five of those images: a straight-on shot, a three-quarter turn, and one with a different expression. That set is your cast member. Everything after this is a photo shoot with her, not a new audition.

Step two: attach the references to every generation

Adding even one reference image switches the generation into edit mode - the model works from the picture rather than from a blank canvas, and the prompt now describes the scene instead of the person. That is the whole mechanism, and it is covered in detail in generating from your own photo.

Practically: keep the same reference set attached to every post, and let the prompt carry only what changes. “Sitting on a balcony at golden hour, linen shirt, holding a coffee” - the face is not in that sentence, because the face is in the references.

Two habits that save real money here. Put the strongest, sharpest portrait first in the reference order, because the result stays closest to the first one. And when a generation comes back wrong, change one thing before you regenerate - it is cheaper to fix the light than to reroll the whole scene.

Step three: batch the month, not the evening

The workflow that survives contact with a content calendar is batching. Sit down once, do fifteen scenes in one pass, and post them over three weeks. This is also where a prompt library stops being a nice idea and becomes the thing that makes Tuesday possible: the same six scene templates, reused with a new location and a new outfit.

Batching also fixes the quality problem nobody talks about. When you generate one image an hour before posting, you post whatever came out. When you generate fifteen and post nine, the feed is the best two thirds of your work.

The rules that decide what you can actually publish

Three constraints, and they matter more than any technique above.

The platform decides, not the generator. Adult output is for you, not for your feed - the same point we made about vertical clips for Reels. If the persona is meant to be public, build her safe-for-work from the first reference image, because a face trained into adult scenes is not a face you can repurpose later.

Disclosure is moving. Platforms increasingly want synthetic media labelled, and the rules change faster than any article can track. Check the current policy where you post. Beyond compliance: an audience that discovers the person is generated, after months of not being told, does not feel clever, it feels handled.

A real face is not raw material. Do not build a persona from photographs of a real person - a model, a friend, someone from a stock site - without their permission and the rights to use them. Our terms require it, and it is the single most common way this kind of project turns into a legal problem.

What you get to keep

Everything the persona produces is yours to use commercially: we take no share and require no credit, which is the part most generators do not allow and which we wrote up separately in selling what you generate. What you do not get is exclusivity - another prompt may land somewhere near your look - so the durable asset is the account, the voice and the consistency, not the face itself.

The persona is built in Fellowi Images; when a still needs to move, Fellowi Video animates it into a vertical clip.