Train an edit finetune from before/after pairs
Last updated: October 15, 2026
Teach Layer one edit from before/after pairs, such as sketch to finished render, and repeat it on new images in your house style.
An edit finetune learns a change rather than a look: each pair shows the image you start from (the before) and the result you want (the after), and the finetune learns to turn one into the other. Use it for edits your team makes over and over, the same way every time: inking a sketch, painting a flat into the house style, turning a daytime scene to night, adding the studio's standard rim light. The pairs live in a reference set, so the same examples stay reusable for the next retrain.
Before you start
- You need a role that can train styles. Ask a workspace admin if you don't see Create finetune.
- Training spends Creative Units. The price shows on the train button before you start; see Creative Units for how runs are charged.
- Put the finished images, the afters, in a reference set first. See Create a reference set.
- One complete pair is enough to start a run. Automatic training needs more pairs; see Trainable models for which models train edits.
Steps
-
Open the reference set that holds the finished images.
-
Under Finetunes, select Create finetune in the Image section, then select Customize.
-
Under Base model, open Select a model and pick an edit model (its name ends in Edit, or is a Kontext model). The grid turns into training pairs: each image in the set becomes the After of a pair, with an empty Before edit slot beside it.
-
Fill each pair's before. Select the empty Before edit slot, pick the starting image from your library and confirm with Select 1 asset. Use Add pair for a pair that isn't in the set yet.
-
If you don't have a before for a pair, select Generate before in its slot. Under Trace the after, Line art, Canny or Depth builds the before from the after's structure; under Prompt edit, a model reverses the edit description on the after. When you have befores but no afters, Generate after works the other way.
-
In Describe the edit under a pair, write the change in one sentence. Then select the copy button in the corner of the field (Apply to all pairs) to give every pair the same description. If other pairs already have their own text, choose Replace all or Fill empty only.
-
Select Create finetune. The run starts in the background and the model's row in the Image section shows Training.
Result
The finetune joins the reference set as a new row under Finetunes, on the edit model you picked, and you get a notification when it's ready. Open the row to see its runs and preview images. To use it, open the row, then choose Forge with model from the Model actions menu beside its name, which opens an image edit with the finetune applied; or add the reference set to an edit, as in Generate with a trained reference set. Give it a new before image and the edit description, and it applies the edit it learned.
Production tips
- Teach one edit per finetune. Every pair should show the same change. Mixing "add color" pairs with "remove background" pairs leaves the finetune guessing which one you want.
- Keep everything else identical inside a pair. Same framing, pose and crop in the before and the after. Anything that differs besides the edit, the finetune learns as part of the edit.
- Vary the subjects across pairs. Different characters, props and scenes teach the edit itself; ten pairs of the same character teach that character.
- Use your real starting material. Pairs made from the sketches your artists actually draw beat traced line art, because the finetune sees the inputs it will get in production. Traced befores are a good way to grow a small set.
- Write the description as you'll prompt it. The finetune learns the edit under that wording, so keep it short, use it on every pair, and prompt the same way when you apply the finetune.
- Grow the set and retrain. Add the best results as new pairs, then retrain. The current version keeps working until the new one is ready; see Judge and improve a finetune.
If something goes wrong
- Create finetune is greyed out with "Give at least one pair a before image to train". No pair is complete yet. Fill a Before edit slot.
- It says "Describe the edit on N pairs to train". Every pair needs a description. Fill the empty ones, or use Apply to all pairs.
- The run failed with "Some training inputs couldn't be used". Select Review pairs, check that every image opens and every pair has both a before and an after, then try again. See Training failed for other failures.
- The finetune changes more than the edit. The pairs differ in more than the edit (crop, pose, lighting). Line them up, remove the odd pairs and retrain.