Review and edit training captions

Last updated: October 15, 2026

Read and rewrite the caption each training image trains with, so the finetune learns your subject and leaves the rest up to the prompt.

Every image in a reference set trains with a caption. Training ties whatever the caption does not mention to the finetune, and keeps whatever it does mention controllable from the prompt. So a caption should describe what changes from image to image (pose, framing, setting, props) and leave out what you want learned. Why this works, and how the trigger word fits in, is in Captions and trigger words.

Before you start

  • Open a finetune, not only the set. Captions show under each image while a finetune is open in the reference set. On the set's overview the images show without them.
  • A role that can train. Editing captions needs edit access to the set; retraining needs the Train styles permission and costs Creative Units (see Creative Units).

Steps

  1. Open the reference set and select its finetune in the Finetunes list. The References tab shows each image with its caption below it.

  2. Read the caption under each image. It is the text that image trains with, written by Layer the first time the set trained.

  3. Click a caption and rewrite it. Name the subject the way the set does and describe only what is particular to this image. Click away to save.

  4. To let Layer write a caption, leave it empty. The box reads Add a caption…, and Layer fills it when the next run starts. It never overwrites a caption you wrote.

  5. Select Retrain finetune. Captions change the finetune only when it trains again; the current version keeps working until the new one is ready.

  6. Open the finetune again to read the caption Layer wrote, and correct it before the next run if it describes something you want learned.

Result

The edited captions are saved on the set and used by every run from now on, for every base model you train from it. The run you started produces a new version on the model's Finetunes tab. Compare it with the previous one in Judge and improve a finetune.

What a good caption says

Character sets

Refer to the character by the set's name and describe what varies: pose, camera angle, framing, expression, setting and props. Leave out the traits that make the character who they are (face, hair, signature outfit) and the art style, so the finetune learns them. Layer writes its own captions this way, and puts the set's name in front of any caption that lacks it.

Style, object and scene sets

Describe the content of each image: the subject, the view and the colours of the things in it. Leave out words for the style itself (the medium, the line, the rendering), since that is what the finetune should learn. Don't describe the background colour of item or icon art; keep those backgrounds consistent instead.

Production tips

  • Use one vocabulary. Call the same thing by the same word in every caption ("staff", not "staff" in one and "walking stick" in another), and use the words you will prompt with later.
  • Describe what you want to swap out. If a character wears the same coat in every image but production needs other outfits, describe the coat in each caption so it stays promptable instead of being learned.
  • Keep captions short. A sentence or two is enough. Case doesn't matter: training reads captions in lowercase.
  • Edit pairs caption the edit. A set of before and after pairs trains on Describe the edit for each pair, not on these captions. See Train an edit finetune from before/after pairs.

If something goes wrong

  • No captions show under the images. Open a finetune from the Finetunes list; the overview doesn't show them.
  • A caption goes back to its old text. The save failed. Edit it again and click away; if it keeps reverting, reload the set.
  • The run stops before training, with a caption error. Layer could not write a caption for an empty one. No Creative Units are spent; write the caption yourself or start the run again.
  • The finetune copies a detail into every image. That detail appears in most images and no caption mentions it. Describe it where it appears, or vary it in the set, and retrain. See Finetune results look overcooked or ignore the prompt.

Related