Train a style finetune
Last updated: October 15, 2026
Train a finetune on your house art style so items, scenes, UI and anything else you generate come out looking like your game.
Before you start
- Training needs a role with the Train styles permission, and costs Creative Units; the Create finetune button shows the cost. See Creative Units.
- Decide what kind of asset the finetune is for. One set per kind (icons, backgrounds, UI) trains a tighter finetune than one set that mixes them. The sections below say what each needs.
What a style set teaches
A finetune learns what its images have in common. For a style, the content should change from image to image and the look should not: the same rendering, line, lighting and palette over many different subjects. Whatever stays the same, including things you did not mean to teach (a frame, a background colour, a camera angle), comes back in every result.
Choose the images
Art style
Pick finished pieces in the final style over a wide range of subjects: characters, props, places, creatures. Leave out anything in another style, early concepts and sketches, and images with UI, watermarks or text unless you want them learned.
Items and icons
Keep the presentation identical: the same view (all front-on, or all three-quarter), the same lighting direction, the same stroke weight and outline, the same palette and level of detail. Put each item on a transparent or neutral background, centred, with similar margins. Vary the items themselves: weapons, potions, food, tools, gems. A finetune trained this way returns new items that slot into the same inventory grid.
Backgrounds and scenes
Use images of a consistent size and orientation, all landscape or all portrait, at the same aspect ratio as the backgrounds you will generate. Vary the content: forest, town, cave, interior, day and night. Keep the camera height and the level of detail consistent if your game uses them consistently.
UI elements
Keep each element the same size in its image and centred, on a transparent or neutral background. Vary the shape and colour: buttons, frames, panels, badges and bars in different colours. If every button in the set is blue, the finetune learns blue as part of the style.
Steps
Create a reference set and add the images, as in Create a reference set.
Name the set for the look and the kind of asset, for example Frostpeak item icons. Keep the category as Art Style for a general look; choose Object for a family of items or Scene for backgrounds, with the chip above the name.
Under Finetunes, click Create finetune on the Image card, choose Customize and pick a starting model. Automatic picks the model and settings for you and skips to the last step.
Check the captions under the image tiles (Add a caption…). Describe the content of each image: what it is, its view, its colours. Keep the style itself out, or to one short phrase: the finetune carries the look, and any style wording you add becomes something your prompts have to repeat. Leave a caption empty and Layer writes one when training starts. See Captions and trigger words.
Add Preview prompts for subjects the set does not show ("a frost lantern icon", "a desert market at dusk") to see whether the look carries over to new content.
Click Create finetune.
Result
The finetune appears under Finetunes on the set, with the images its preview prompts rendered, and applies whenever the set is used on the model it trained on. Prompt the content and let the finetune carry the look. See Generate with a trained reference set.
Production tips
- Grow a small set by generating. Start with the pieces you have, train, then generate the subjects the set lacks. Keep only results you would ship, add them to the set and use Retrain. A few rounds turn a handful of pieces into a set that covers your whole asset list. The current finetune keeps working until the new version is ready.
- Keep prompts about content. Words that name another style or medium ("watercolour", "photorealistic") fight the finetune.
- Put fixed wording in Always applied. For presentation every asset needs ("game item icon, centred, transparent background"), open the set's Prompts tab and fill Always applied. The text joins every prompt that uses the set; it is not part of training, so you can change it at any time without retraining.
- One style per run. Combine a style finetune with character or object sets, not with a second style finetune; two styles average each other out.
- Batch the asset list in a workflow. A workflow that names the style set can render a whole icon sheet or a run of backgrounds in one go. See Find and run a workflow.
- Clean up after generating. For icons and UI on a transparent background, remove the background from the results before they go into the engine.
If something goes wrong
- Every result has the same composition, or a background you did not ask for. The set repeats something it should vary. See Finetune results look overcooked or ignore the prompt.
- The look is faint. The finetune is undertrained or the weight is low; the same article covers that case.
- The training run failed. See A training run failed.