Finetunes

A finetune is a model trained on the files in one of your reference sets, so a base model learns your character, style, edit, voice or motion directly. It is one thing a reference set can do: the set stays the source of truth, and the finetune is the strongest way to apply it on the model it was trained for.

How it works

A reference set already works on every model that can read it: Layer passes its files, description, palette and Always applied text to whatever the model accepts (see Reference sets: the source of truth for your art pipeline). Training goes one step further. It takes the set's files and teaches one base model to reproduce the concept, and saves the result in the set as a finetune.

When you later attach the set to that model, Layer applies the finetune first, with a weight control for its strength. On any other model, the set falls back to its files and text. A set can hold finetunes for several models at once, each used where it applies.

Each training run adds a version on the set's Finetunes tab. One version per base model is active; Use this one switches to another, and Retrain makes a new version while the current one keeps working until the new one is ready. Every run, finished or not, is listed under Runs.

When to train

Attach the set without training

Most work starts here. Attaching an untrained set gives a model your reference images and wording, which is enough for exploration, for models that take several reference images, and for concepts that are still changing. It costs nothing extra and works across models.

Attach a one-off reference

For a single image that should borrow a pose, a palette or a composition, attach the image to that generation. Nothing is saved for the team, and nothing is learned beyond that run.

Train a finetune

Train when you need the tightest consistency, at volume, on one model you use heavily: a hero character across hundreds of poses, a house style across a whole content drop, the same edit applied to a whole asset library. Training costs Creative Units and takes time, and the finetune only works on its own base model, so it pays off when the concept is settled and the output volume is high.

Kinds of finetune

Which kinds you can train depends on what the set holds and on the base model you pick. Trainable models lists the models for each kind.

Image finetune

A LoRA (a small adapter added to an image model) trained on single images. It learns what the images share: a character's identity, an art style, a family of items, a type of scene. You use it in text-to-image generation with the set attached. It needs a minimum number of images, and consistent images matter more than many. See Train an image finetune.

Edit finetune

A LoRA trained on before-and-after pairs instead of single images. It learns a transformation: what changes between the before and the after. You use it when you edit an image with that edit model and the set attached. See Train an edit finetune from before/after pairs.

Image finetune or edit finetune

The two are easy to confuse, because both are LoRAs trained from images.

• What it learns. An image finetune learns a subject or a look, and draws it from a prompt. An edit finetune learns a change, and applies it to an image you give it: turn a sketch into a finished render, add your house outline to any icon, swap a character into its winter costume.
• The dataset. An image finetune trains on single images, each with a caption describing it. An edit finetune trains on pairs, each with an editing caption describing the edit (Describe the edit under the pair). The before and after should match in everything except the change.
• Where you use it. An image finetune runs when you generate from a prompt on its model. An edit finetune runs when you edit an existing image on its edit model.

Automatic training can train both from one set: an image finetune from the images, and an edit finetune when the set holds enough complete pairs.

Image style

Some image models train a style of their own from a small set of reference images, with no LoRA and no weights file. The style stays on that model, and you choose how closely generations follow it (Match: Precise or Flexible). Use it for a brand or illustration look you want on that model's output, raster or vector. See Train a Recraft image style.

Voice

A custom voice for speech models, trained from audio clips in a set. Clone copies a voice from reference clips; Design creates one from a written description. You then speak any script in that voice. See Train a custom voice and Create speech and voice-over.

Motion

A 3D motion saved in a set, made from a text description or extracted from one reference video. Any rigged character can then play it. See Turn a video or text into a reusable motion and Rig and animate a 3D character.

Imported finetune

A LoRA you trained elsewhere, brought in as a .safetensors file with Upload finetune. You choose its base model and enter its trigger word, and the set then uses it like a finetune trained in Layer. See Import or download finetune weights.

What it means for you

• Curate the set before you train. A finetune copies what the set holds, including its mistakes; see How training works for why the dataset matters more than the settings.
• Captions decide what the finetune ties to your trigger word or character name. See Captions and trigger words.
• A finetune is tied to one base model. If your team moves to another model, train again on it; the set and its files carry over.
• Training needs a role with the Train styles permission, and downloading weights needs Download style weights.
• Training spends Creative Units. See Training cost and time.
• The agent can set up a set, draft its captions and start training for you. See Train a finetune with the agent.

Related

• Train an image finetune
• Captions and trigger words
• How training works
• Generate with a trained reference set
• Reference sets: the source of truth for your art pipeline