Training makes a new avatar from photographs you supply. Once it finishes, that character is available to generate with like any other avatar.
Before you start
You need a paid plan, and the subscription has to be currently active. A plan that has lapsed, is past due, or is still on a trial will not do — training is expensive to run, so it is gated on a subscription that is actually in good standing. The wizard tells you if your account does not qualify.
You need photographs of one subject. How many is set by us and shown in the wizard — enough that the model sees the subject from more than one angle, and capped so a set does not become unmanageable.
What makes a good set:
- One subject only. Other people in shot teach the model the wrong thing.
- Variety in angle, expression, lighting and background. Twenty near-identical photos teach less than eight different ones.
- Clear and sharp. Blurred, tiny or heavily filtered pictures produce a blurred, unreliable avatar.
- Consistent identity. A decade of changing appearance averages into somebody who does not exist.
Consent is required, explicitly
You confirm you have the right to train on these photographs before anything starts, and that confirmation is recorded with the request.
This is not a formality. If the photographs are of another person, you need their agreement — training builds something that can generate new pictures of them indefinitely, which is a bigger thing to agree to than one photograph being used once. Do not train on someone who has not agreed.
The terms in the app are the full statement of what is permitted.
Running it
Open the training page — the avatar menu is not shown to everyone yet, so it may not be in your sidebar, but the page works.
Give the avatar a name, add your photographs, confirm consent, and start.
Then leave. Training takes a while and needs nothing from you. It moves through several stages, which the page reports as it goes:
| Stage | What is happening |
|---|---|
| Uploading | your photographs are being received |
| Validating | they are being checked as a usable set |
| Captioning | each photograph is being described automatically, so the training has something to learn against |
| Training | the actual work, and the longest part |
| Creating avatar | the result is being turned into an avatar you can use |
| Completed | ready to generate with |
If something goes wrong the request is marked failed and says so. You can cancel one that is still running.
When it finishes
The new avatar appears in your list. Use it exactly like any other — see avatars.
Expect to test it before relying on it. Generate a handful of images across different scenes and lighting to see where it holds up and where it drifts. That is normal, and it tells you what kind of prompt suits it.
If the result is disappointing
Almost always the photographs rather than the settings.
- Too few, or too similar → add variety and train again
- Subject unclear in several shots → remove the weak ones; a small clean set beats a large noisy one
- More than one person in some shots → remove them
- Heavy filters or make-up varying wildly → pick a more consistent set
Retraining with a better set is the fix. There is no adjustment to a finished avatar that recovers a poor dataset.