How to create an AI influencer with a face that stays the same

Making a beautiful AI photo is easy. Anyone can do it in thirty seconds. The trouble starts with the second photo, when you notice it isn't the same person: a different nose, a different jaw, a different body. A character who changes face in every post isn't an influencer. It's a collection of strangers.

This guide is what we learned building fanlygen, and getting it wrong many times before it worked. It's written for people who want to understand how it works, even if they end up doing it by hand with other tools.

Why the face changes

Image generators start from zero on every request. Ask for "brunette, 25, green eyes" and you get one. Ask again tomorrow and you get another woman who fits the same description but is not the same person. Words can't pin down a face: "brunette with green eyes" describes millions of different faces.

Consistency comes from one thing only: showing the model what that specific person looks like, with images, not adjectives. That step is called training, and in practice it's done with a technique called a LoRA: a small add-on to the image model that has learned one person.

The four steps, in order

1. The face

You start from a portrait. You can generate it from a description (this is the one moment where words are the right tool, because you're choosing a face, not reproducing one) or start from a photo you have the right to use, such as your own.

The most useful advice in this whole article: pick a face with something recognisable. A mole, a particular smile, a dimple. Perfectly regular faces are harder to keep stable, because there's nothing for the model to hold on to.

2. The body, front and back

This is the step almost everyone skips, and it's the one that ruins the result later. With only a portrait, the model knows the face and invents everything else. Ask for her standing, from behind, or in a different pose, and you get an average body, not hers. From one photo to the next her build changes, and people notice even when the face is right.

You need at least two full-body references, front and back, and they have to actually show the body. A reference in a loose dress tells the model how wide the dress is, not what she looks like.

3. The photo set and the training

From the face and body you generate a set of images of the same person in different situations, lights and framings, usually around fifteen. It's study material: it teaches the model that she stays herself in a different light, half-length or far away, smiling or serious.

The training runs on that set. It's the slow part (tens of minutes) and the part that really costs money: generating images costs cents, training costs a few euros. Knowing that changes how you work. Generate dozens of photos without thinking twice, but train only once you're sure about the character.

4. After that

Once training is done you have something new: a model that knows how to make that person. From then on you put her wherever you want and she stays herself. That's where the real work starts, which is posting regularly.

Five reasons the face still drifts after training

Training fixes most of it, not all of it. When people tell us "she looks similar, but it's not her", it's almost always one of these.

1. The model never saw her body. The face is right but the build changes, usually thinner than she should be. The references were portraits or clothed. Cutting out the background doesn't help: the clothes are on her, and the model keeps copying them.

2. Words beat images, and the words are old. You wrote "curtain bangs" when you created her, then swapped the references for photos without bangs. The bangs keep showing up, because that word is still in the prompt and weighs more than the new images. The rule: if you have identity photos, don't also describe her features in text. Photos or words, never both.

3. The starting photo is of someone else. Recreating a pose from another person's photo brings that person's face along, and for the model she's the main subject: sharp, centred, right there. You have to say explicitly that the identity of the starting photo must be discarded completely: that person is only a mannequin for pose, framing and light. Without that sentence you get her, in the pose you wanted.

4. No second pass on the face. In a full-body shot the face is a few pixels wide and generators are imprecise there. The fix is a dedicated second pass that redraws the head from the reference. Two details matter more than they seem: replace the head, not just the face (hair is half of recognition), and crop the reference to the head, or the pass drags in pieces of what she wore in the reference.

5. A monotonous training set. Fifteen front-facing photos, same light, same distance, teach the model to make that one photo. Step outside that framing and the face falls apart. Vary light, distance and angle; keep only the person the same.

When the face is wrong, the temptation is to change everything at once. That's the best way to never learn what worked. Change one thing at a time, in the order above.

What an AI influencer generator should do for you

If you're comparing tools, these are the questions worth asking, because they're the ones that decide whether you'll be throwing away nine generations out of ten:

• Does it build the body references, front and back, or only the face?
• Does it stop adding written traits once there are identity photos?
• When you start from someone else's photo, does it discard that person's identity?
• Is there an automatic pass on the face, with the head cropped?
• Can you see the character before paying for training?

On fanlygen these steps are already part of the flow: the references are generated full-body, written traits switch off when there are photos, and the face pass is automatic. But the reasoning holds with any tool: the problem is rarely the generator, it's what you're handing it.

How long, and how much

You can build the character in an afternoon, training included. The real costs are two: a few cents per image and a few euros for the training. Anyone who tells you it's free is moving the cost somewhere else, usually onto your time.

New fanlygen accounts get 300 free credits when they confirm their email, enough to start building the character and see her before deciding to train. That's exactly the point where you find out whether she convinces you.

And then?

The model is only the raw material. An AI influencer works for the same reasons a real one does: she posts often, has a recognisable personality, and answers. The difference is that production doesn't cost you a day of shooting. We told the story of one profile that grew once production stopped being the bottleneck in the Carima case study, and if you're thinking about monetising, start with running an AI model on Fanvue.