Train an image finetune

Last updated: October 15, 2026

Train a finetune on a reference set's images so every generation on that model holds your character or style.

A finetune is a LoRA learned from the images in a reference set. The set stays the source of truth: train it once, retrain it as the art moves, and every tool that takes the set applies the finetune on the model it was trained on. What a finetune is, and when to train one instead of attaching images, is in Finetunes.

Before you start

  • A reference set with your images. Build one first with Create a reference set. Image finetunes train on the set's image files only, at a fixed training resolution, so use sources at least that large (see Training settings). If the set holds too few, Create finetune says how many more it needs.
  • A role that can train. Training needs the Train styles permission. Without it the set shows no finetunes.
  • Creative Units. The cost shows on the Create finetune button before you start. Layer reserves it when the run starts and bills only the steps that ran; a run that fails or that you stop releases the rest. See Creative Units and Training cost and time.

Steps

  1. Open the reference set you want to train. Its Finetunes list sits beside the reference files, with one section per kind of file the set holds.

  2. In the Image section, select Create finetune. The form opens on Automatic, which picks the base model and settings for your images. To train that way, go straight to step 6.

  3. Select Customize, then choose the model to train under Base model. The Trainer card below shows which engine trains it; where a model has more than one trainer, pick it there.

  4. Select Show advanced settings and type a Trigger word: one made-up word, such as frontierwarden, that the finetune ties to your subject. Leave Epochs, Learning rate and LoRA rank at their defaults for a first run; Training settings explains each one.

  5. Under Preview prompts, type a prompt you will want the finetune to handle, and use Add prompt for more. Layer renders them as the finetune trains so you can judge the result. They don't change the trained weights.

  6. Select Create finetune. Layer confirms Training started. and the model's row in the list reads Training.

  7. Open the row to follow the run: Training in progress shows how far it has got and the time left. When it finishes, the Finetunes tab holds the new version with its preview images.

Result

The set gains a finished finetune for the base model you trained, shown as the Active version on that model's Finetunes tab and as Ready to use or its settings in the set's list. Layer sends a notification when the run finishes or fails, so you can leave the page while it trains. Any tool that takes the set now applies the finetune on that model: see Generate with a trained reference set.

Layer adds the trigger word to every caption when it trains, and to your prompt when you generate with the set, so you rarely need to type it yourself.

Production tips

  • Curate before you train. Consistency counts for more than quantity: one subject or one style per set, similar framing, and no blurry or upscaled images. A smaller set of strong images beats a large mixed one.
  • Vary what should stay promptable. The finetune learns whatever stays the same across the images. For a character, vary poses, angles, expressions and backgrounds so only the character is fixed; for a style, vary the subjects.
  • Read the captions before the first run. Each image trains with its own caption, which Layer writes for any image that has none. Describe what changes from image to image and leave out what should be learned. See Review and edit training captions.
  • Write preview prompts that test the brief. Ask for what production will need: a new outfit, a new setting, a close-up, a pose none of the images show. A finetune that only reproduces the training images is memorising.
  • Start on Automatic, then customize. Train a first version with the defaults, judge it, and only then change one setting at a time. Judge and improve a finetune covers what to change.
  • Grow the set from its own output. Generate with the finetune, add the best on-model results to the set, and retrain. Each retrain adds a version; the current one keeps working until the new one is ready.
  • One set, several models. A set can hold a finetune for each base model. Train the models your pipeline uses and each tool picks the one that applies.

If something goes wrong

  • Create finetune is greyed out. Hover it to read why. It may need more images, a model picked under Customize, or more Creative Units ("Not enough creative units to train this model. Click to top up."). Clicking the button in that last state opens the top-up.
  • Training failed. The row shows Failed with Open to see why. Open it for the reason and its fix, such as Adjust settings after a timeout or Top up. Creative Units for a failed run are released. See Training failed.
  • The run reads Taking longer than expected. It may have stalled. Select Stop training on the banner and start it again.
  • Start another training run? appears. That model already has a run in progress. Starting another trains a second version and spends Creative Units again; cancel unless you meant to.

Related