Train a character finetune

Last updated: October 15, 2026

Train a finetune on one character so it stays on-model in any pose, outfit or scene you prompt, across a whole production.

Before you start

  • Training needs a role with the Train styles permission.
  • Training costs Creative Units; the Create finetune button shows the cost before you start. See Creative Units.
  • Have finished, approved images of the character ready: renders, turnarounds, key art, expression sheets. The training form tells you when a set has too few files; for sizes and resolution, see Training settings.

Choose the images

A character finetune learns what stays the same across your images and ties it to the set's name. Everything that changes from image to image stays something you can prompt. So the set should hold the character fixed and vary everything else.

Shots

Mix headshots and close-ups (they carry the face and expression) with full-body images (they carry proportions, costume and silhouette). A set of only portraits gives a finetune that cannot draw the legs; a set of only small full-body figures loses the face.

Angles

Show the character from the front, in profile, in three-quarter view and from the back. A turnaround sheet is a good source: crop each view into its own image rather than training on the sheet.

Expressions and poses

Include a range of expressions and poses, not ten copies of the same neutral stance. A finetune trained on one pose returns that pose whatever you ask for.

Backgrounds

Vary the backgrounds: some plain, some in scenes, in different colours and lighting. If every image has the same backdrop, the finetune learns it as part of the character and paints it in behind every result.

Consistency

Every image must show the same design: the same costume details, colours, markings and proportions. One off-model piece (an early concept, a fan redraw, an alternate outfit) pulls the whole finetune towards an average. Keep alternate outfits in a set of their own.

Quality

Use sharp, full-resolution images where the character fills most of the frame. Leave out blurry or upscaled images, and crop away UI, watermarks, signatures and other characters.

Steps

  1. Create a reference set and add the character's images, as in Create a reference set.

  2. Name the set after the character, for example Captain Rhea, and click the Art Style chip above the name to change the category to Character. The name is the character's anchor: training ties the character to it, and your prompts use it to call the character up.

  3. Under Finetunes, click Create finetune on the Image card. Choose Customize and pick a starting model, so you can check captions and set preview prompts. Automatic picks the model and settings for you and skips to the last step. The set's name already works as the trigger, so leave Trigger word under Show advanced settings empty.

  4. Check the captions under the image tiles (Add a caption…). Leave a caption empty and Layer writes its caption when training starts; for a Character set, that caption names the character by the set's name and describes only what varies. If you write your own, do the same: use the name, and describe the pose, expression, framing, setting and props, never the hair, face, costume or art style. See Captions and trigger words.

  5. Add Preview prompts that test what the set does not show: a pose it never strikes, an environment it never stands in, a framing it never uses (for example "Captain Rhea climbing a rope ladder in a storm, seen from below"). They render while the finetune trains and do not change it.

  6. Click Create finetune. Training runs in the background; you get a notification when it finishes.

Result

The finetune appears under Finetunes on the set, with the images its preview prompts rendered. It is applied whenever the set is used on the model it trained on. Mention the set with @ to place the character in the sentence ("@Captain Rhea kneels by the fire"); if you leave the mention out, the name is added to the start of the prompt for you. See Generate with a trained reference set.

Production tips

  • Test for generalisation, not likeness. A finetune that reproduces the training images is easy; one that draws the character in a new pose, a new place and a new framing is what you need. Judge it on the preview prompts that ask for something the set never shows.
  • Prompt what varies, not who it is. At generation, describe the action, setting, framing and outfit changes. Describing the character's fixed traits again ("red hair, eye patch") is unnecessary and can pull against the finetune when your wording differs from the design.
  • Put fixed wording in Always applied. For presentation your pipeline always needs ("full body, plain grey background"), open the set's Prompts tab and fill Always applied. The text joins every prompt that uses the set; it is not part of training, so you can change it at any time without retraining.
  • Change outfits through the prompt. A set with varied backgrounds and poses but one costume still accepts "in winter armour". Keep a separate set when an outfit must be exact.
  • Grow the set from your best outputs. Generate the angles and expressions the set lacks, keep only the on-model results, add them and use Retrain. The current finetune keeps working until the new one is ready.
  • Combine it with a style. Use the character set together with a trained or untrained style set to render the character in a house look; each set does its own job.
  • Use it in workflows. A workflow that names the character set produces on-model shots whoever runs it. See Find and run a workflow.

If something goes wrong

Related