Use trained models and reusable assets

Reuse trained weights and saved identities from your Library.

Updated 2026-10-22

Wavemaker does not train models for customers. You can't prepare a dataset or start a training job from a conversation, the web app or the command-line tool. What you can do is keep a character, product, place or brand consistent with references and saved Library identities, and reuse trained weights that are already registered in your workspace.

Training isn't available

• The agent in a conversation has no training step. Asking it to train a LoRA or fine-tune a model does not start one.
• Studio and its training tools are staff-only. If you open them, you see Studio isn't available with the message "Everything Studio did now happens in a conversation. Describe what you want to make and Wavemaker sets it up." and a Start a new chat button.
• Registering new weights from the web app is also staff-only, so the trained weights you see in Library are ones already registered in your workspace.

Keep a subject consistent without training

Most reasons to train a model (the same face in every shot, the same bottle and label, the same room) are handled with references.

1. Upload clear photos of the subject with Upload files, or choose a saved identity with Add from Library.
2. In the brief, say what the subject is and which features must not change, for example "keep the label, cap color and bottle shape exactly as in the photo."
3. For a person, product or place that appears in several shots, the agent usually makes an identity plate first: a still of the subject that later shots are made from. Check it before asking for more shots.
4. Check a small first result for faces, proportions, clothing and product details before asking for more shots.

See Use references, characters and styles for how to give each reference a role.

Save an identity to reuse

To keep a result as a reusable identity, open the conversation's Outputs view, open the ••• menu (Conversation options) and choose Save to Library. This saves the conversation's selected output; if nothing is selected yet, Wavemaker shows "Select an output first, then save it to Library."

The form asks for:

Choose Save asset. Wavemaker shows "Private asset version saved." and opens Library. A saved identity is private and can't be changed, and work that already used it stays pinned to it. Save to Library always saves version 1, so give each save a new Asset identifier; reusing an identifier that already holds different content is refused with "Asset versions are immutable; choose a new version". Saved identities appear in Library under Characters, subjects & voices. See Reuse an exact Library item.

Trained weights in your workspace

Registered weights appear in Library as a Trained weights group, described as "Registered weights in this workspace." Each entry is a named model at one version, with the base model it was made for. A version never changes once registered; a later version is a separate entry.

For each version's status, open Library tools and choose Trained weights. Each card there shows Registered model · version N, a status and a Run and output status line:

When a newer version of the same model exists, its card in My library says "Version N is available. This item stays pinned to version M." with a Review newer versions link. Choose the version you mean; Wavemaker never switches to the newest one for you.

On that card, choose Inspect exact version to open the item. When the weights are stored, the item page offers Download exact source, which downloads the .safetensors file.

If the weights came from Wavemaker's earlier training service, the dataset and training run behind them show as unavailable. The weights themselves still work.

Use trained weights in a conversation

1. In the composer, choose Add from Library. The Choose from My library window opens.
2. Set the Kind filter to trained weights, or search for the model's name, and select the exact version.
3. Give it a purpose, for example "lead character", then choose Add to conversation. In a new chat that hasn't started yet, the button reads Start chat with 1 instead.
4. In the brief, say what the weights are for, for example "use my character weights for the lead in every shot."
5. Send the brief.

You can also start from Library: open the item and choose Use in a conversation. Both routes need a workspace role that can write.

Adding weights to a conversation does not generate anything or spend credits. The conversation then shows a Registered model record card for that version.

What happens when a run uses your weights

When the agent runs a graph that names your model, Wavemaker checks that your workspace registered that exact version and still holds its bytes, checks again before the job starts, then loads a copy into that job's isolated workspace. Your weights are not retrained, and other workspaces can't use them. The run is charged in credits like any other graph run; see Credits.

Weights run only on a runtime that accepts weights for the base model they were registered for, up to that runtime's size limit. Weights for one base model can't be used with another.

If a run with your weights fails

Test before relying on it

Run one small test on a prompt or scene the model was not trained on, then compare faces, proportions, products and style against your references. A model loading without an error does not show that it gives useful results. If the test misses, change the references or the brief before asking for more shots.

Use references, characters and styles · Reuse an exact Library item · Check graph and method requirements

Related guidance

• Understand credits, reservations and charges
• Check graph and method requirements
• Create and edit an image
• Library: find, organize and reuse your work
• Use references, characters and styles