Generate with a trained reference set

Last updated: October 12, 2026

Generate with a reference set you trained as a finetune, so every image holds your character or style at the strength you choose.

A finetune is a LoRA (or, on some models, a trained custom style) learned from your set's images. Creating the set and starting training are covered in Create a reference set; how a set reaches each kind of model is in Reference sets: the source of truth for your art pipeline. This article is about using the result.

Before you start

The set has a finished finetune. Open the set and look at Finetunes. A finished one shows its settings or Ready to use; while it trains the row reads Training. Until it finishes, runs use the set's images and text instead.
A finetune works only on the model it was trained on. A set can hold finetunes for several models and uses each one where it applies. Hosted models that take no custom weights, such as Gemini, never apply a finetune; on them the set falls back to its images and description.
Training needs a role with the Train styles permission. Anyone who can generate can use a set that is already trained.
Generating with a finetune costs Creative Units like any other run. See Creative Units.

Steps

In a session, open Create image from the side panel.
Click Reference set and select your trained set. Sets with a finetune carry a Finetune badge; turn on Only finetuned to list only those.
Click Add 1 reference set. If the model you are on cannot apply the finetune, Layer switches to the one that can and says so: "Good news! [model] applies this reference set as a trained LoRA, so we switched to it." Click Keep followed by your previous model's name to switch back. The set now sits under LoRA, with its weight beside it.
Describe the scene in the prompt: what happens, where, and in what framing. The finetune carries the look, so you do not need to describe it.
Set the weight in the box beside the set, or drag its slider. It starts at 100% and runs from 10% to 200%. Lower lets the prompt lead; higher holds closer to the set.
Click Create image.

Result

The images land on the session's board. Open one and check Reference sets in its details: a set that applied as trained weights reads Used as finetune.

How Layer picks the model

When you add a set, Layer compares every model in the tool for how directly it applies the selection. It switches only to a stronger match: a model that applies the set as trained weights beats one that uses only its images, and one that uses its images beats one that only reads its description. It switches when you add or remove a set, never when you pick a model yourself.

If you then choose a model that applies the set less directly, an orange notice above the prompt says so and names the model that applies it, with a Use button to switch. In the model menu, Recommended for this reference set lists the models that apply it best.

Other ways to start

From the Reference Sets page

Right-click a trained set and choose Use as LoRA. A new session opens on the model the set is best applied by, with the set attached. Use as Reference opens the default image model instead and uses the set's images. A set with no finetune offers a single Use. The same choices appear when you right-click a set on a session's board.

From the set itself

Click Use in session in the set's header. The arrow beside it lists the set's finished finetunes under Use a model; picking one opens a session on the model that finetune was trained on.

From the LoRA button

On a model that applies LoRAs, the form has a LoRA button beside Reference sets. Its picker starts on sets with a finetune.

With the agent

Type @ in the chat box and pick the set. The agent uses it for the generations it runs. See Write a brief for the agent.

How the prompt and the finetune work together

A LoRA is trained on the set's name. When the set applies as a LoRA, an @ mention of the set in your prompt is replaced with the name it was trained with, so you can place it in the sentence ("@Frontier Warden kneels by the fire"). A Character set you did not mention gets its name added to the start of the prompt. Renaming the set later is safe: the finetune keeps the name it learned.

In the side panel a set applying as a LoRA shows under LoRA with its weight, not as a chip in the prompt.

Production tips

Find the weight once, then reuse it. Fix the Seed under Advanced and run the same prompt at a few weights, for example 60%, 100% and 140%. Note the one that holds the look without freezing the pose or composition, and use it for the batch.
Keep the prompt about content. Words that name another style or medium ("watercolour", "photorealistic") fight the finetune. Describe the subject, action, setting and framing.
Combine one trained style with other sets. A trained style plus an untrained character or object set each do their own job. Two style finetunes in one run average each other out.
Batch it in a workflow. Generation nodes take reference sets with the same weight control, so a workflow can render a whole sheet of shots with the finetune applied. See Find and run a workflow.
Retrain as the art moves. Add approved new pieces to the set and use Retrain; the current finetune keeps working until the new one is ready, and Use this one rolls back to an earlier version.
Similarity is the image's own weight. Where a model weighs an attached image, that image carries its own weight in the form, which sets how closely the result follows it. It is separate from the finetune's weight.
Edit finetunes edit. A finetune trained on before-and-after pairs applies on the edit model it was trained on: attach the image to change as the Source image and describe the change.

If something goes wrong

There is no weight beside the set, or the model did not switch. The set is not applying as a finetune on this model: its training has not finished, it was trained for a different model, or the model takes no custom weights. Check Finetunes on the set, and see Results don't follow your reference set for every cause.
The result does not look like the set. Start with the weight and the prompt, then work through Results don't follow your reference set.
A run with a LoRA and an Image prompt fails. Remove the image from Image prompt and run again. To turn an image into prompt text instead, select it on the board while the tool is open and click Describe image beside the prompt.
Training failed. The model's row under Finetunes shows Failed and Open to see why. Open it to read the reason and the fix it offers: Top up, Adjust settings, Review pairs or Get help. A finetune trained earlier keeps working while you retry.
There is no Create finetune button. Your role cannot train, or the set has too few images. Ask a workspace admin, or add images until the card reads Ready to train.