Models (your account)

This is the account-wide Models page, headed Model Training Studio. There is also a Models section inside each project, which turns trained models on for that project.

A custom model is trained on your own files so it reliably reproduces one character, style, look, motion or voice. This page is where you build the files into a dataset, start a training run and watch it work.

• Get started — the kinds of model you can train.
• Trained models — your models, finished or still training.
• Train a new model — the Train new model form.
• Datasets — the sets of pictures, clips or voice recordings you train on.
• Training jobs — every training run, with its progress. Logs shows a run's log, and Hide logs closes it.

1. Under Datasets, press New dataset, choose its type, give it a name and press Create dataset.
2. Add files with Upload, or press From project to pick ones already in a project and then Add to dataset.
3. Open the dataset to caption each file, then press Save once it has enough. That makes it ready.
4. Under Train a new model, pick the ready Dataset and a Base model, name the model, pick the project under Charge to project, and press Train model.
5. Watch it under Training jobs. When it finishes, turn it on in a project's Models page.

A dataset's own page has Upload images, or Upload clips for clip and voice sets. Auto-caption with AI writes picture captions for you. Save captions, or Save transcripts for a voice set, keeps your edits.

Freeze & ready makes the dataset ready to train. Back to Model Training Studio returns to this page. Delete dataset removes a dataset and its file list; trained models keep working and the files are kept.

Each trained model has Rename and Archive. Archive takes it out of the model picker without changing past generations, and archived models don't count toward your plan's limit. Show archived lists them again, with Unarchive.

Continue training, on a finished model that can resume, trains it further. You choose how many more steps, a new version is made, and you pay only for those steps.

The form also has Steps, LoRA rank (net dim), Learning rate and, for a clip set, Motion length: how many frames of each clip it learns from.

Training details opens how a model was trained: its Recipe, Base model, Base version, Steps, LoRA rank, Learning rate, Resolution, Dataset, Trigger word and when it was made.

Training is billed on the render time it actually uses. The form shows an Estimated cost before you start. If your plan has no room for another model, it says so, with Manage plan.

Cancel training, or Cancel under Training jobs, stops a run. You pay for the render time used so far and get no model. Once a checkpoint is saved, Stop & keep checkpoint stops it and keeps a usable, partial model for that cost.