Skip to content

Running nodes & jobs

Press Run on a node and it executes against the values currently on its input ports — so run upstream nodes first (or run the chain from the top) so their outputs exist. Fast nodes return inline; generation nodes hand off to the job system. A full-graph RUN (top bar) first shows the run’s estimated cost, node by node, before dispatching anything.

Image, video, and audio generations run as jobs on background workers:

  • Progress is shown on the node while the job runs.
  • Jobs are resumable — refresh the page, close the laptop, come back: a running job re-attaches and completes into the node.
  • Long renders are fine. Video jobs are given generous completion windows; a slow model won’t be abandoned mid-render.

Results land in two places at once: inline on the node, and in the project’s assets.

Every run is recorded on its node. Open a node’s generations to compare earlier results, bring one back, or download it. The prompt, model, and provider that produced each result are stored with it.

  • Change one thing at a time — a prompt tweak, a model swap — and re-run just the affected node.
  • Downstream nodes keep their last output until you re-run them, so you can regenerate step 2 of 5 without losing steps 3–5.
  • For batch nodes like Collection, you can regenerate individual images from the batch instead of the whole set.

A failed run shows the error on the node — most commonly a missing input (run the upstream node first) or a model-side rejection. Fix and re-run; failed attempts don’t overwrite your last good result.