Train a finetune with the agent

Last updated: October 15, 2026

Hand the agent your images and it builds the reference set, drafts the captions, and trains a finetune once you pick the base model.

The Layer agent can run a whole training for you from chat. It makes a reference set from the images you attach, writes a caption for each image and shows them to you, recommends a base model, and starts the run once you choose. It stops at the two points where your judgment matters: the captions, and the model you spend Creative Units on. The set it makes is an ordinary reference set, so you can open it, add images and retrain it later from the reference set editor.

Before you start

  • You need a role that can train styles. Without it, the agent can still create the reference set and review captions with you, and says so when you ask it to train.
  • Training spends Creative Units once you pick a model; see Creative Units.
  • Have the images ready in your library or on your computer. A finetune needs several consistent images of one subject; see Train a character finetune for what makes a good set.

Steps

  1. On Home, select Add to prompt, then Attach files or photos, and pick the images to train on.

  2. Ask the agent to train a finetune on them. Say what it is (a character, an art style, an object or a scene) and what to call it, then send the message. The agent opens a session, creates the reference set and shows its drafts in a Review captions card.

  3. Read each caption. A good caption describes what changes from image to image (pose, framing, setting) and leaves out what should stay the same, because the finetune learns what the captions leave out. Leave a caption on Accept, choose Edit to rewrite it, or Regenerate to have the agent draft it again. Then select Submit.

  4. The agent recommends a base model in a card. Pick the one you want and select Submit answers. Your pick is the go-ahead: it starts the run and spends Creative Units. Select Cancel instead and nothing trains; the agent waits until you pick.

  5. Training runs in the background. Ask the agent how it's going at any time, and it checks the run and tells you how far along it is.

Result

The finetune trains on the reference set the agent made, and you get a notification when it's ready. Ask the agent to try it, and it generates a few test images with the finetune for you to judge; or generate with the set yourself, as in Generate with a trained reference set. The set, its captions and its runs are in the reference set editor like any other.

Production tips

  • Give the mode and the name up front. "A character finetune called Frontier Warden" saves the agent a question. For a character, the set's name is what the captions use to refer to the character, so choose the name you'll prompt with.
  • Spend your time on the captions. They decide what the finetune can be told to change later. Edit any caption that names something you want to keep fixed, or misses something you want to vary. See Captions and trigger words.
  • Name a model if you have a house default. Name it in your message and the agent uses it if it can be trained; if it can't, the agent says so and offers one that can.
  • Edit the set in chat. Ask the agent to add or remove images, rename the set or change its category. New images need captions, so the agent reviews them with you again before a retrain.
  • Switch to the editor for fine control. For trainer settings, preview prompts or edit pairs, open the reference set and train it there; see Train an image finetune.

If something goes wrong

  • The agent says training isn't available to you. Your role can't train styles in this workspace. Ask a workspace admin for access; the agent can still prepare the set and its captions.
  • The agent says there aren't enough images, or captions are missing. Attach more images of the same subject, or review the captions it lists, and ask it to try again.
  • The model you asked for isn't offered. Not every model can be trained. The agent proposes one that can; Trainable models lists them.
  • The run failed. Ask the agent what happened: it reads the run and tells you the reason. See Training failed for what to do next.

Related