Import or download finetune weights

Last updated: October 15, 2026

Bring a LoRA trained outside Layer into a reference set, or download a finetune's weights to use in your own tools.

A finetune's weights are a single .safetensors file. Importing one adds it to a reference set as a version for the base model it was trained on, so the set applies it wherever it applies that model, beside the set's images and text. Downloading gives you the file Layer trained, to run in your own pipeline.

Before you start

  • The weights file. One .safetensors LoRA per import. You can upload it from the picker, or add it to the Asset Library first.
  • The base model it was trained on. Layer can import weights only for a model it can train, and a LoRA applies only to the model it was trained for. The models are listed in Trainable models.
  • Its trigger word. An import needs the word the LoRA was trained with.
  • Permissions. Importing needs the Train styles permission. Downloading needs Download style weights; without it the download button doesn't show.

Import a LoRA

  1. Open the reference set the weights belong to.

  2. Select Upload finetune (the upload button beside Create finetune in the Image section). Pick the .safetensors file, or use Upload to add it from your computer, then select Select 1 asset.

  3. In Import a trained model, check the Base model is the one the LoRA was trained on and type its Trigger word. Under Parameters used to train the model, enter the values it was trained with if you know them; they are a record and don't change the weights.

  4. Select Import model. The weights join the set as a new version of that base model, and its row in the Finetunes list shows the parameters you entered.

Download a finetune's weights

  1. Select the finetune in the set's Finetunes list and open the Finetunes tab. On the version you want, select Download weights. Your browser saves its .safetensors file.

Download weights is also in the column header of each version under Compare.

Result

An imported LoRA is a version on its base model's Finetunes tab, like a trained one, and you choose which version is active there (see Judge and improve a finetune). It has no preview images, because Layer did not train it. A downloaded file is the LoRA Layer trained for that version, for the base model shown on its row.

Production tips

  • Use the trigger word it was trained with. Layer adds it to prompts when it applies the finetune, so a wrong word weakens the LoRA without any error.
  • Test an import before you rely on it. Generate a few images with the set on the imported version, and compare them with what the LoRA gave you in your own tools at the same weight.
  • Keep the set as the source of truth. Put the images the LoRA was trained on in the set too. You can then retrain it in Layer later, and anyone using the set sees what it is meant to look like.
  • Name downloads before you share them. Rename each file with the set, base model and version, for example frontier-warden_flux-dev_v2.safetensors, before they go into a shared drive or an engine project.
  • Check the licence. Weights trained on a base model carry that model's licence terms. Check them before you ship or redistribute a downloaded or imported LoRA.

If something goes wrong

  • The picker shows no weights file. The picker lists only .safetensors files. Use Upload in the picker to add yours.
  • "No base model is available for training in this workspace." The workspace has no model it can import for. Ask a workspace admin, or contact support.
  • The import fails. The dialog shows the reason. Check that the file is a LoRA for the base model you chose, not a full model checkpoint or a LoRA for another model.
  • Generations ignore the imported LoRA. It applies only on its base model, and only when it is the active version. See Results don't follow your reference set.

Related