Judge and improve a finetune

Last updated: October 15, 2026

Compare a finetune's versions by their preview images, choose the one your generations use, and know what to change before you retrain.

Every training run adds a version to the base model's Finetunes tab, and exactly one version per base model is active: the one every tool applies when it uses the set on that model. A retrain becomes the active version when it finishes, so check it against the one before and switch back if it is worse.

Before you start

  • Preview prompts. Each version shows the images rendered from the set's Preview prompts, the same prompts for every version, which is what makes them comparable. Set them up when you train: see Train an image finetune.
  • A role that can train. Switching versions and retraining need the Train styles permission.

Steps

  1. Open the reference set, select the finetune in its Finetunes list, then open the Finetunes tab. Each version shows its preview images, newest first, and the one in use reads Active version.

  2. Select Compare. The versions line up in columns, with one row per preview prompt.

  3. Use 2× or 4× on a row to zoom in. The whole row zooms and pans together, so you see the same detail on every version. The expand button on an image opens it full size.

  4. Select Make active on the version that holds the subject best. It becomes the active version straight away.

  5. Select Versions to go back to the list. The version you chose reads Active version; the others keep Use this one so you can switch again at any time.

Result

Every tool, workflow and agent run that uses the set on this base model applies the active version from now on. The other versions stay on the tab until you delete them.

What to look for

Is it on-model?

Check the features that define the subject in every preview: face, silhouette, colours, costume details, or for a style the line, palette and rendering. A preview that changes one of them is drifting. See A trained character drifts off-model.

Does it follow the prompt?

A preview that ignores what the prompt asked for (a new outfit, a setting, a pose) and returns something close to a training image is overcooked: it learned the images rather than the subject. See Finetune results look overcooked or ignore the prompt.

Does it hold up in real work?

Preview images cover only a handful of prompts. Before you commit, generate a few images from your own briefs with the set attached and try a lower and higher weight. See Generate with a trained reference set.

Improve it

Change one thing per retrain, so the comparison tells you what helped. Each retrain is a new version; select Retrain finetune on the model, and the current version keeps working until the new one is ready.

Curate the images

Remove images that are off-model, blurry or repeat the same pose, and add images that cover what the previews got wrong. Good generations from the finetune itself, picked by an art director, are a quick way to grow a small set.

Fix the captions

A detail that sticks to every result is usually in every image and in no caption. Describe it, or vary it. See Review and edit training captions.

Adjust the training settings

Under Show advanced settings, fewer epochs or a lower learning rate ease an overcooked finetune; more epochs help one that barely changes the base model. What each setting does is in Training settings.

Production tips

  • Keep the preview prompts fixed. Changing them between runs leaves you comparing different prompts, not different versions. Add prompts rather than replacing them.
  • Write prompts that stress the finetune. Include the shots production needs most and the ones the training images don't show: a close-up, a back view, a new setting.
  • Record why a version is active. When a team shares a set, note the reason in the set's description so nobody switches back to a version that was rejected.
  • Delete versions you will not use. Delete version on the version list removes it from the set for good. Creative Units already spent on it aren't returned.
  • Score at volume. For a batch made with the finetune, an output scoring rule can check every image for you before hand-off. See Set up scoring rules for a workspace or project.

If something goes wrong

  • There is no Compare. It appears once the model has more than one version.
  • A version has no preview images. It was trained with no preview prompts, or it was imported from weights trained elsewhere. Generate with it to judge it.
  • Make active or Use this one does nothing. Layer shows "Couldn't switch which version this set uses." Try again; if it persists, reload the set.
  • Every version has the same problem. The cause is in the set, not the settings. Review the images and captions before retraining again.

Related