Avatars: Locking a Character So It Looks the Same Tomorrow

Describe a person in words and you get a different person every time. “A woman with dark hair in her thirties” is not a character — it is a category, and the model picks a new member of it on every generation.

For a single image that does not matter. For anything that runs — a channel posting five times a week, a brand face, a serial drama — it is fatal. An audience will forgive almost any production shortcut except not recognising who they are looking at.

An avatar is the fix. It is a character the models already know, so you get the same person repeatedly across as many shots as you like. This page is the mechanics: where avatars come from, how to train one that holds up, and how to generate with it. For why identity persistence is the asset a channel is actually built on, and what quietly destroys it months later, see building a channel persona.

Why this is the feature the business rests on

On TJP this is the central capability rather than an accessory — the platform’s own framing is consistent AI talent, with the identity locked and everything else bending around it: hair, eyes, expression, outfit, pose, props and style stay controllable while facial and body likeness hold. The contrast it draws with other tools is exactly the failure described above — a character that mutates shot to shot.

Three consequences follow, and they are why this feature and not raw image quality is the thing to evaluate a platform on:

Without it you have clips. Individually fine, collectively incoherent. There is no series, because there is no continuing subject.
With it you have a cast — and a cast is a format. The same character in a new situation every week is most of what short-form drama and most brand channels actually are.
It is what makes volume worth anything. Once the character is fixed, the marginal cost of the next episode is the shots, not the casting. That is the basis of producing a content calendar in an afternoon instead of a quarter.

There is a practical gate too: the flagship image model does not run without a connected avatar. Selecting it without one prompts you to connect one. So an avatar is not a refinement you get to later — it is what unlocks the best output on the platform. The video tools, img2img, editing, upscaling and the partner models all work without one.

Where an avatar comes from

Two routes, with very different requirements:

Route	What it needs	When it fits
Claim	A connection to an existing avatar	Fastest way in; no training, no wait
Train	Your own photographs, an active paid plan, and consent	You need a character nobody else is using

Some avatars are also available to everyone and need no claim at all.

If you are still deciding whether this route is for you, start with one of those or a claimed one. A trained avatar is a commitment of money and time, and it is the wrong thing to spend on before you know your format works — see what an episode actually costs.

Training your own, properly

Training builds a new avatar from photographs you supply. Done well it is the strongest asset in the operation. Done carelessly it produces a character that looks slightly different in every shot — which is precisely the problem you were trying to solve, now paid for.

What it requires before you start:

A paid plan, currently active. Not lapsed, not past due, not a trial. Training is expensive to run and is gated on a subscription in good standing; on TJP this is the top tier.
Photographs of one subject. The wizard states how many.
Recorded consent, confirmed before anything begins.

What makes a good photo set — this is where results are won or lost:

One subject only. Other people in shot teach the model the wrong thing.
Variety in angle, expression, lighting and background. Twenty near-identical photographs teach less than eight different ones. This is the most common mistake by a distance.
Clear and sharp. Blurred, small or heavily filtered pictures produce a blurred, unreliable avatar. A filter you like is a filter the model learns as part of the face.
Consistent identity. A decade of changing appearance averages into somebody who does not exist. Pick a period and stay inside it.

Training runs on its own — uploading, validating, then the training itself — and needs nothing from you while it goes. Start it and leave.

Keep the training set. Archive the photographs and the settings you used somewhere off the platform. If you ever lose account access, that archive is the only thing that lets the persona be rebuilt.

If the photographs are of another person, you need their agreement, and you should have it in writing.

The reason is worth stating plainly: training does not use a photograph once. It builds something that can generate new pictures of that person indefinitely, in situations they never saw and cannot anticipate. That is substantially more than consenting to one image being used, and treating the two as equivalent is how people end up in disputes they cannot win.

Do not train on someone who has not agreed. The platform’s terms are the full statement of what is permitted there; rights, likeness and consent covers how to keep this defensible once money is involved.

Generating with one

Simpler than working without one, and the prompts get shorter.

Pick the avatar in the panel, then write the prompt as usual — describing the scene, the lighting and the treatment, but not the person.

standing at a rainy bus stop at night, red wool coat,
neon shop signs behind her, cinematic film still

The avatar supplies who. Your prompt supplies everything else. This is why prompting improves once you are working with one: you stop spending words on a description that never quite lands and spend them on what you actually care about. Everything in prompting for video still applies, minus the subject.

Adding the person back into the prompt fights the avatar rather than helping it. If you find yourself writing “a woman with dark hair”, delete it.

Presets come with avatars — ready-made looks and settings that suit that character. Apply one as a starting point and change what you like; a preset sets the controls, it does not lock them.

Continuity the avatar does not cover

An avatar fixes identity. It does not fix everything an audience reads as “the same person in the same story”, and the rest is on you:

Wardrobe goes in every prompt containing the character. The avatar does not remember the red coat.
Location is better handled by generating one approved still of the space and animating from that still repeatedly, rather than re-describing the room and getting a slightly different one.
Time of day and weather drift silently unless restated.
Seed and model held constant across a run keep the look related — and a seed is not transferable between models, so a mid-run model switch resets it.

The working habit that removes most of this: keep a short character sheet — a paragraph of wardrobe and physical signature — and paste it into every prompt for that character. It costs nothing and eliminates an entire class of continuity error.

What an avatar is not

Not a licence to generate a specific real person. What you may make, and of whom, is governed by the platform’s terms and by the consent you actually hold. Likeness use is permitted where it is owned, licensed or approved — not merely where it is technically possible.
Not a guarantee of a perfect match every time. It is a very strong constraint, not a photocopier. Shots still need reviewing.
Not permanent by itself. Model versions move underneath a persona and will change how it renders. That risk, and the monthly check that catches it, are covered in building a channel persona.