Train a finetune for a cast of characters

Last updated: October 15, 2026

Train one finetune on a whole cast that shares a look, so new characters you describe come out in that look and fit the roster.

Before you start

  • Training needs a role with the Train styles permission, and costs Creative Units; the Create finetune button shows the cost. See Creative Units.
  • Decide what you want back. This article is for a shared look across many characters: a roster, a family of enemies, a set of NPCs. To keep one character exactly on-model, train one set per character instead (see below).

A cast set or one set per character

A finetune learns what its images have in common. Train it on many different characters drawn the same way, and what they share is the look: proportions, line, rendering, palette, how faces and costumes are built. That is what you get back, applied to whatever character you describe.

It does not learn any one character's identity well. Each character appears in only a few images, and the finetune averages them. Use one character finetune per character when:

  • a character recurs in many assets and must be recognisable every time;
  • exact details (markings, costume, a logo on a shirt) must hold;
  • you want to call the character up by name in a prompt.

The two work together. Attach a character set and the cast set to the same run, and the character holds while the cast set keeps the house look.

Choose the images

Variety of characters

Include as many different characters as you can: different builds, ages, species, silhouettes and costumes. The more they differ in content, the more clearly the shared look stands out, and the better the finetune handles a character it has never seen.

One consistent format

Present every character the same way: the same framing (all full body, or all bust), a similar pose or view, the same background treatment and a similar size in the frame. Format is part of what the finetune learns, so a consistent one comes back consistently; a mix of turnarounds, action shots and portraits gives a mix back.

One look only

Every image should be in the final house style. Leave out early concepts, sketches and pieces from another project, even if they show a character you need.

Steps

  1. Create a reference set and add the cast images, as in Create a reference set.

  2. Name the set for the look, for example Frostpeak cast, and leave the category as Art Style. Do not choose Character: that category treats every image as the same character and ties them all to one name.

  3. Under Finetunes, click Create finetune on the Image card, choose Customize and pick a starting model.

  4. Check the captions under the image tiles (Add a caption…). Each caption should describe its character: build, face, hair, costume, colours, pose and view. Describing each one is what keeps identity something you prompt, rather than something the finetune decides. Leave a caption empty and Layer writes one in that form when training starts. See Captions and trigger words.

  5. Add Preview prompts for characters that are not in the set ("a stout dwarf blacksmith with a braided beard, full body"), in the set's format. They show whether the look carries over to someone new.

  6. Click Create finetune.

Result

The finetune appears under Finetunes on the set and applies whenever the set is used on the model it trained on. Prompt a new character by describing them; the finetune supplies the house look. See Generate with a trained reference set.

Production tips

  • Write prompts the way the captions read. The finetune learned descriptions of characters in a certain format; prompts in the same shape (who, build, costume, colours, pose, view) land closest to it.
  • Grow the roster. Generate new characters, keep the ones that fit the look, add them and use Retrain. Each round adds variety, which makes the finetune more general.
  • Split off the stars. When a generated character becomes a regular, collect their best images into a Character set and train it on its own.
  • Keep a style set for everything else. If you also need props and environments in the house look, train those in their own sets; a cast set learns characters in a character format. See Train a style finetune.

If something goes wrong

Related