Train a custom model

Why train a product

An untrained product still generates, using its uploaded photos as reference, but the result is an approximation. A trained model has learned the product's exact shape, texture, and detail, so it holds up across many different shots without drifting.

How it works

1. 1

   Upload source images

   At least one image is required to start training, though a handful of clear photos from different angles gives a noticeably better result than a single shot.

2. 2

   Start training

   Training runs in the background and usually takes 15 to 30 minutes.

3. 3

   Automatic validation

   Once training finishes, Misu generates a set of validation images with the new model and checks them against your source photos. The product only locks once that check clears the quality bar; otherwise it's marked failed and can be retried.

4. 4

   Generate

   A locked product is ready to use anywhere in Generate, Video Studio, Campaigns, or Canvas.

Cost and failures

Training a product spends tokens once, at the start, not per generation afterward. If the training run fails to start for any reason, those tokens are refunded automatically.

See also