A trained character drifts off-model
Last updated: October 15, 2026
A character finetune applies, but the character comes out with the wrong face, costume, colours or proportions, or looks like someone else.
First open a result and check Reference sets in its details. If the set does not read Used as finetune, the finetune did not apply, and the result followed only the set's images or text; see Results don't follow your reference set.
Most likely causes
The references are inconsistent
The finetune learns what the images share. Early concepts, redraws by different artists, alternate outfits or older versions of the design teach an average, and the character comes out as a mix.
Fix: keep only images of the current, approved design, all with the same costume details, colours and proportions. Move alternate outfits into a set of their own. Retrain.
The set holds more than one character
A Character set treats every image as the same character. A group shot, a second character in the frame or images of a sibling character blend into one identity.
Fix: crop or remove other characters so each image shows only this one. For a roster, train one set per character, or a cast set in the Art Style category; see Train a finetune for a cast of characters.
The captions describe fixed traits
When a caption spells out the hair, face, costume or colours, the finetune ties those traits to the words instead of to the character's name. Prompts that do not repeat the exact words then lose them.
Fix: open the training form and check the captions under the image tiles. Name the character and describe only what varies: pose, expression, framing, setting, props. Retrain. See Captions and trigger words.
The category isn't Character
A set trained as Art Style learns the character as a look, with captions that describe its appearance and not its name, so identity holds loosely and the background is learned with it.
Fix: click the category chip above the set's name, choose Character, and retrain.
The prompt fights the character
A prompt that describes the character's traits differently from the design ("blonde" for a red-haired hero), names another style, or asks for a second person in the frame pulls the result away from the finetune.
Fix: mention the set with @ where the character belongs in the sentence and describe the action, setting and framing, not the character's appearance. If you remove the mention, the name is still added to the start of the prompt, so you do not need to type it twice.
The weight is too low
At a low weight the prompt leads and the model's own idea of the character takes over. Raise the set's weight in the generation form and compare at a fixed Seed. If the character only holds at a weight that also freezes pose and background, the finetune is overcooked; see Finetune results look overcooked or ignore the prompt.
The set lacks the view you asked for
A character trained only from the front drifts in profile or from behind; one trained on portraits drifts at full body. The finetune has to guess what it never saw.
Fix: add images of the missing angles and framings (generate them, keep only the on-model ones) and retrain. See Train a character finetune.
Still stuck?
Contact support from Help & Resources > Talk To Us in Layer. Include the reference set's name, the model it was trained on, the prompt and weight you used, and a result that shows the drift. Contact support with the right details lists everything that helps.