Run batch inference

You need: the curator dataset role or higher.

Batch inference runs segmentation across a whole dataset in the background, instead of image by image.

Steps

1. Open the dataset's inference page.
2. Assign a model per label. You are not restricted to one model for the whole dataset — different classes can be served by different models.
3. Choose the write mode — see below.
4. Start the run, and watch the progress panel. The work is queued to a Celery worker, so you can leave the page.

Per-label model planner → write mode → start → progress panel running.

Patch vs. replace

Mode	What it does
Patch	adds the run's predictions alongside what is already there
Replace	removes existing annotations in scope and writes the run's output instead

Replace discards existing annotations

Including corrected ones. If a review pass has already happened on this dataset, replacing throws that work away. The UI warns before it does this — read the warning rather than clicking through it.

Ordering follows the label hierarchy

The work list is ordered by your label hierarchy, so parents are segmented before the children searched for inside them. If your hierarchy is wrong, the run order will be too.

Next: Review and correct