Finetune results look overcooked or ignore the prompt
Last updated: October 15, 2026
A finetune applies, but every result looks the same, comes out harsh or burnt, or ignores what you ask for; or the opposite, the look is too faint to see.
Is it overcooked or undercooked?
An overcooked (overfit) finetune has memorised its training images instead of learning the look. Signs:
- every result has the same pose, composition or camera angle;
- a background, prop or frame from the training images appears when you did not ask for it;
- results copy a training image almost exactly;
- colours are oversaturated, edges crunchy, details noisy;
- the prompt changes little: a new setting, action or outfit is ignored.
An undercooked (underfit) finetune has not learned enough: results look generic, the style is faint, or a character is only roughly like the set.
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 at all; see Results don't follow your reference set.
Fix an overcooked finetune
Lower the weight
The quickest fix, with no retraining. Lower the set's weight in the generation form and run the same prompt again with a fixed Seed under Advanced, so you compare like with like. Step down until the prompt leads again while the look holds.
Use an earlier version
A training run can save its finetune more than once on the way, and each save appears under Finetunes on the set as a version with its epochs and the previews it rendered. An earlier one has seen the images fewer times and is often more flexible. Compare the previews and click Use this one on the version that follows your preview prompts best.
Train for less
Retrain with Customize and, under Show advanced settings, lower Epochs or Learning rate. Change one at a time so you can tell which helped. A lower LoRA rank also limits how much the finetune can memorise. See Training settings.
Vary what should stay promptable
A finetune learns whatever every image shares. If all images have the same background, pose or framing, so will every result. Add images that vary those things (new poses, settings, backgrounds, compositions), drop near-duplicates, and retrain.
Caption what varies
Anything the captions leave out is learned as part of the set. If the captions never mention the background, the background becomes part of the look. Describe in each caption what changes in that image (the pose, setting, background, framing, props), so those things stay tied to words you can change. See Captions and trigger words.
Fix an undercooked finetune
Raise the weight
Raise the set's weight in the generation form and compare at a fixed seed. If you have to push it far up to see the look, retrain instead.
Train for more
Retrain with Customize and raise Epochs under Show advanced settings. Higher epoch counts generally pair with a lower learning rate, so when you raise Epochs, leave Learning rate as it is or lower it.
Make the set more consistent
A set that mixes looks teaches an average of them, which reads as weak. Remove pieces in another style, early concepts and off-model images, keep one concept per set, and retrain. Consistency matters more than count.
Keep the prompt about content
Prompt words that name another style or medium ("watercolour", "photorealistic") pull against the finetune and make it look weaker than it is.
Test before you batch
Add Preview prompts when you train that ask for things the set does not show: a new subject, setting or framing. They render at each saved version, so you can see where a run starts to overcook before you pick which version to use.
Still stuck?
Contact support from Help & Resources > Talk To Us in Layer. Include the reference set's name, the model it was trained on, a prompt that shows the problem, and the weight you used. Contact support with the right details lists everything that helps.