Why the face drifts between shots
Text-to-image generation does not remember earlier results. The phrase "a woman with chestnut hair" gives a new person every time: similar to the description, but never the same one. The shorter the description, the wider the spread.
To make a face repeat, the model needs a visual sample. You can supply it on the fly by attaching photos to the request, or in advance by training dedicated weights for one hero.
Method 1: reference photos in the request
In photo editing mode you upload one, two or three photos of the hero and write an instruction: what to change, which setting to show. The model takes facial features from the source images and carries them into the new frame. Nothing needs training, and each frame costs the usual price.
- good for a short series and one-off tasks;
- the more varied the source photos (angles, light), the steadier the result;
- likeness holds less well than with a trained character, especially at difficult angles.
Method 2: train your own character
You upload 4 to 30 photos of one person or one fictional hero, choose the number of steps and the trainer type. After a while the service returns a character you can pick when creating a picture. Face, hairstyle and overall style hold noticeably better than with references.
The recommended number of photos is about 15. The service does not accept fewer than four, and more than thirty adds no quality. There are two trainers: a general one (fits people, products and characters) and a portrait one (tuned for faces and better at keeping features).
| Steps | General trainer | Portrait trainer | What you get |
|---|---|---|---|
| 1000 | $2.60 | $3.12 | recognizable, not perfect |
| 2000 | $5.20 | $6.24 | the sensible middle, most common choice |
| 3000 | $7.80 | $9.36 | maximum likeness, takes longer |
Which method to choose
| References in the request | Own character | |
|---|---|---|
| Photos needed | 1-3 | 4-30, about 15 is best |
| Upfront cost | none | from $2.60 for training |
| Price per frame | $0.06-$0.15 (by class and resolution) | $0.05 |
| Face stability | medium | high |
| When to pick it | one-off tasks, a quick test of an idea | a series of dozens of frames, one hero across a project |
A simple rule: if you will make fewer than ten frames, start with references. If the hero must be drawn constantly, training pays for itself within the first few dozen pictures.
Which photos to upload
- the same person in every shot, with nobody else in the frame;
- different angles: front, three quarters, profile;
- even light without harsh shadows, face in focus;
- varied clothes and backgrounds, so the model does not memorize them as part of the hero;
- no heavy filters or masks, decent image size rather than tiny thumbnails.
Upload only your own photos, photos of people who gave explicit consent, or images of a fictional hero. Photos of strangers, acquaintances and public figures cannot be used without permission.