Results don't follow your reference set
Last updated: October 12, 2026
You attached a reference set, the generation ran, and the result does not look like your style, character or object. Usually the set was applied only in part, or not at all, by the model you used. Find what you see below.
Check what the set contributed
Open the result and look at Reference sets in its details. Each set shows how it reached the model: Used as finetune, Used followed by a number of reference images, Used in prompt, or Didn't contribute. That tells you which cause below applies. A run where a set did not contribute still completes and is still charged, so check this before a large batch.
On the form, before you run, hover a set's tile to see how the chosen model will apply it: Applied as a trained finetune (LoRA), Applied as image references, Applied through the prompt only, or This model can't use the reference set.
The model can't apply the set well
The set won't be applied to this model
An orange warning on the form says the set won't be applied to your model, and Didn't contribute shows on the result. The model takes neither a finetune from this set nor its images, so nothing from it reached the model.
Fix: click the Use button in the warning to switch to the model it names. In the model menu, models are grouped as Recommended for this reference set and Partial match; models that can't use it are under Show incompatible models. If you picked a model by hand, let Layer choose: when you attach a set, the form switches to the model that applies it most directly and tells you so, with a button to keep your old choice.
Only the prompt carried the set
can only approximate this reference set via the prompt, so results may be inconsistent, or Used in prompt on the result. The model could not read the set's files, so it got the set's description as text. Text alone rarely holds a look or a character.
Fix: use the model the warning recommends. If you have to stay on this model, write the set's description as the prompt words you would use for the look, and add exact colours under Color palette; both reach every model.
The finetune was trained on a different model
A trained finetune works only on the model it was trained for. On any other model the set falls back to its images or its description, and the result notes A finetune is available on a different base model. Switch to that model, or train the set for the model you want to use.
Only some of the set's images were used
The tooltip says Uses a number of a larger number assets. Each model has room for a limited number of reference images, and some take only one per set. Layer fills that room from the start of the set, so the first image matters most.
Fix: put the clearest, most representative image first in the set. If the model only has room for a few, pick a model that takes a finetune or more references, or train the set.
The set was removed when you changed model
A message such as can't use camera movements, so followed by the set's name is no longer attached means the set's category means nothing to the new model (a camera move on an image model, a character on a sound effect). Layer leaves it out. Attach the set to a model of the kind it is for.
The set applied, but the result still drifts
The finetune is too weak or too strong
When a set applies as a trained finetune, its tile shows a weight. Drag it to change it, or double-click the value to type one. It runs from 10% to 200% and starts at 100%. Raise it towards 150% when the style is there but faint; lower it towards 50% when the style overpowers the subject and every result has the same composition. The weight applies only to finetunes, not to sets used as images or text.
Training isn't finished, or failed
A set is used as a finetune only once its training shows Trained. While it is Training in progress, Starting up or Finalizing…, and if it shows Failed or Cancelled, runs use its images and text instead. Wait for training to finish, or start it again from the set's Finetunes tab. See Create a reference set.
The prompt pulls against the set
A prompt that names another style, artist, medium or look ("watercolour", "photorealistic") competes with the set, and the prompt often wins. Describe what to show, not how it should look, and let the set carry the look.
Two sets are pulling in different directions
A style set and a character set work together, each doing its own job. Two style sets average each other out. Attach one style at a time, or make a new set that holds the combined look.
The set itself is inconsistent
The model learns what the set's files have in common. A set that mixes styles, includes off-model pieces, or holds a character, an environment and a UI kit together teaches a blurred average.
Fix: keep only images you would ship, all in the same look. Split mixed sets into one concept each, and set the category to match: a character trained as Art Style learns its background as part of the character. Then retrain. See Reference sets: the source of truth for your art pipeline for what makes a good set.
Still stuck?
Contact support from Help & Resources > Talk To Us in Layer. Include the session link, the reference set's name, the model you used, and what the Reference sets section of the result says. Contact support with the right details lists everything that helps.