You have written the prompt fourteen ways. You have tried "same woman," "consistent character," "photorealistic," a locked seed, a reference to the last image, and a paragraph describing her cheekbones like you are filing a police report.

She still shows up with a different nose.

Every creative working with AI people hits this wall and concludes they are bad at prompting. They are not. The prompt was never the bottleneck.

The model cannot tell you from Tuesday

When a system learns a specific person, it is not learning a person. It has no concept of bone structure. It does not know that a face is a three dimensional object that stays the same when the light moves. It has a set of images, and it looks for what they have in common. That is the entire mechanism.

So say every reference photo of Sara came from the same afternoon. Same window light, same three quarter angle, same soft smile. Nine photographs. Genuinely nice ones.

The model has no way to know which of those things is Sara.

Is Sara the cheekbones, or is Sara the warm light falling across the cheekbones? Is Sara that jawline, or is she the angle you happened to be standing at? The model cannot ask you. It learns "Sara" as one fused bundle: face, light, angle, expression, mood, all welded into a single thing.

Then you ask for Sara in profile, under hard studio light, not smiling.

You get a stranger. Or you get Sara with golden light glued to her face in a room that has none. Or you get something in between that your eye rejects a half second before your brain can explain why.

Variance is what teaches the model where you end and the room begins.

Photographs of the same person under different light, from different angles, with different expressions, hand it the one thing it actually needs: a chance to notice what stays the same. What survives all that change is the person. Everything else was circumstance.

That is the whole principle. The rest of this is consequences.

Where reference sets quietly fail

• Angle collapse. If every reference is shot from the front, the model has no information about the side of a head, so it invents one. The invention is statistically reasonable and it is not that person. Ask for a profile and you get a plausible stranger wearing her hair. Cover front, both three quarters, and true profile. The profile is the one everyone skips, because nobody likes their own profile. It is also the one that cannot be guessed.
• Lighting bake-in. When the light never changes across your references, the model cannot separate the light from the face, so it learns them as one unit. You have not built a person. You have built a person shaped lighting setup, and it will fight you the first time you ask for night. Soft and directional at minimum. You are not looking for good lighting. You are looking for different lighting, because the difference is the lesson.
• Expression lock. References that are all neutral produce a model that has never seen this face move. Ask for a real laugh and you get the uncanny one, mouth doing something the eyes have not agreed to. Muscles pull a face in ways specific to that face. Give it neutral, a real smile, and something in between. Expression range is not vanity coverage. It is the difference between a person and a mask.
• The filter trap. This is the one that surprises people, because filtered selfies look better to us. A beauty filter does not just soften skin. It moves geometry. Eyes slightly larger, jaw slightly narrower, nose slightly shorter, and it does this differently in every frame depending on angle and distance from the lens. So you have handed the model images of several slightly different people and asked it to find what they have in common. What it finds is a blur. Meanwhile the smoothing strips out pore texture, and the model learns that this person has plastic for skin. No prompt puts it back, because as far as the model knows, it was never there.

More photos is not the answer

The reflex, once you understand the above, is to throw volume at it. Forty images instead of nine.

But forty frames from the same shoot are not forty examples. They are one example, photographed forty times. Same light, same room, same day, same energy. You have not added information. You have added emphasis, which is worse than nothing, because now the model is very confident about a person who only exists in that room.

Coverage is the unit that matters, not count. A small set that deliberately varies angle, light, and expression will outperform a large set that varies none of them.

Which means you are not collecting your best photos. You are collecting the honest ones.

Why we build profiles instead of asking for selfies

Full disclosure on the bias here. This is what Trueth does, and it is why the intake looks the way it does.

We do not ask talent for five nice pictures. We ask for a specific set, shot with specified coverage, unfiltered, and we send them back when they are not. That is not fussiness. Every requirement on the list is one of the failures above, closed before it can happen.

A profile is not a folder of images. It is a person, described completely enough that a model can find them. And it is built with them rather than around them: they applied, they agreed, they get paid every time their likeness generates something.

The useful part for you is that the work holds up when you ask it to do something specific. Not because the tool is smarter. Because the reference was better.

The takeaway

If your AI people fall apart when you move the camera or change the light, stop rewriting the prompt. Go look at what you gave it.

Nine photos from one afternoon will only ever make one person, in one room, on one afternoon. If that is what you need, you are done. If you need someone who can walk into any brief you have this quarter, the model needs to have seen them survive a change.

The prompt is the last five percent. The reference is everything before it.