Why the model forgets the hero
The cause is not a bad prompt. The cause is that every generation is independent.
The model does not remember the previous shot. For it, every run is a blank page. When you write «a 30-year-old man, short beard, grey jacket», you are describing not a person but a whole class of people: millions of faces fit that description. The model picks from that class again each time, and it is completely logical that it picks differently.
Text simply lacks the precision needed. Even if you describe the shape of the eyes, the nose and a mole on the cheek, you are still describing a type of appearance, not a specific face. So fighting the problem at the prompt level is pointless, however many adjectives you add.
There is one solution: stop describing the character in words and start showing it with a picture.
Method 1. A reference image
The basic and most effective technique. You upload an image of the character as a sample, and the model relies on it.
An important nuance: a reference is not the same as a starting frame. The model animates a starting frame, building motion from it. It uses a reference as a source of appearance and builds the scene from your description. The second is more flexible: you can put the same person in another location, other clothes, other lighting.
What makes a reference good:
- The face is large and in focus. If the hero takes up a tenth of the frame, the model has nothing to read the features from.
- Neutral lighting. The model will take hard shadows and coloured lights as part of the appearance and carry them forward.
- A neutral expression. An ear-to-ear grin on the reference risks sticking to the character forever.
- Several samples beat one. If the model accepts more than one reference, give it different angles.
Method 2. A character sheet
The logical development of the first method, and what separates an amateur from someone who does this at volume.
A character sheet is a set of images of one hero from different sides: front, profile, three-quarter, full length, close-up. It is assembled once and then used in every scene.
Why it matters. One front-facing reference works great while the hero looks into the camera. As soon as they turn sideways, the model has to invent a profile it has never seen, and it invents it differently every time. Give it a profile in advance and there will be nothing to invent.
How to build a character sheet in practice:
- Generate or take one good front-facing image of the hero.
- Run it through an editing model (for example, Nano Banana Pro), asking for the same person in profile, then three-quarter, then full length. Editing models are exactly the ones that can change the angle while keeping the face.
- Discard shots where the face changed. That is normal, there will be rejects.
- Save the final set. This is your hero, and from here on you use only that.
Method 3. Last-frame chaining
A technique for long scenes where the hero must pass continuously from clip to clip.
The idea: you generate the first segment, take the last frame of the finished clip and feed it as the starting frame of the next one. The model starts the new scene exactly where the previous one ended and carries over not only the face but also the clothes, lighting and setting.
The plus: the join comes out almost seamless, and the character cannot change because the model builds from a real frame rather than a description.
The minus: errors accumulate. If the hero’s face drifted a little in the fifth clip, the sixth inherits the already damaged version, and the seventh makes it worse. So the chain cannot be stretched forever: every few scenes go back to the original character sheet to reset the accumulated drift.
Method 4. Prompt discipline
It does not replace references, but without it they work worse.
Describe the character once and then copy that description word for word. Not «a man in a jacket» in one scene and «a guy in a grey jacket» in another. The model reads any change of wording as an instruction to change something.
Do not describe the appearance again if you gave a reference. This is the main mistake. When the model has both a picture and a detailed text description of the face, they start to conflict, and the model looks for a compromise between them. You end up with a person who resembles both the reference and the description but equals neither. If there is a reference, describe the action and the scene in the prompt, not the hero.
Separate the character from the context. «The same man, now in the kitchen, pouring coffee, morning light from the window». The hero comes from the reference, the rest from the text.
Which method to use when
| Method | What it solves | When to use | Weak spot |
|---|---|---|---|
| Reference | Sets the face instead of text | Always, it is the base | Works worse at unusual angles |
| Character sheet | Covers every angle of the hero | A series of 3 or more scenes | Needs to be built once, it takes time |
| Last-frame chaining | Continuity between scenes | Long edited sequences | Errors accumulate from scene to scene |
| Prompt discipline | Removes the conflict of text and picture | Always, together with a reference | Does not give consistency on its own |
The right set is almost always this: a character sheet as the foundation, references in every scene, prompt discipline on top, and last-frame chaining only where you need a continuous transition.
A step-by-step workflow for a series of shots
Here is what it looks like in practice if you need a clip of eight scenes with one hero.
- Build the hero. Generate a front-facing portrait and refine it until you like it. Do not spare attempts here: you will see this face many more times.
- Build the character sheet. Use an editing model to get the profile, three-quarter and full length. Discard the failures.
- Debug the prompt on a cheap model at low resolution. You are checking not the picture but whether the model holds the face. You do not need resolution for that, and the price difference is fivefold.
- Generate the scenes. In each, feed the references and describe only the action. A model that accepts several samples at once wins a lot here.
- Pull long stretches through last-frame chaining, but every three or four scenes go back to the character sheet to reset the drift.
- Regenerate the final at high resolution with the already debugged prompts.
Five mistakes that make the face drift anyway
- A low-quality reference. A small face, noise, blur. The model reads what it sees, and what it sees is mush.
- Describing the appearance when there is a reference. The text conflicts with the picture, the model looks for a compromise, you get a third person.
- Asking for an angle that is not in the references. Asking for a view from behind without giving a single profile? The model will invent the back of the head itself, and a new one each time.
- An endless last-frame chain. Errors pile up, and by the tenth scene the hero has become their own distant relative.
- Changing the wording between scenes. Every rewrite of the description is a signal to the model to change something.
How to apply this to a virtual blogger: how to create an AI influencer.
Bottom line
Character consistency is not a property of the model and not the result of a lucky prompt. It is a process: first you fix the hero as a set of images, then you rely on them everywhere, and in text you describe only what the hero does.
Models that claim character stability between shots and accept several references at once save a lot of time here. But even they will not help if you describe the face in words every time and hope for a match. Prompt fundamentals are covered in the article on how to write prompts for video.
