Runs.

A run evaluates work in your graph to produce results. It might create an image, transform an existing file, return text, or call another service. You can run your working graph before saving it as a named project. Running and saving are separate actions.

This guide covers Studio's graph canvas. For programmatic execution, use API runs, progress and cancellation.

#Choose what to run

In the desktop Graph view, a node's run control (tooltip: Evaluate node) targets that instance. Run Graph in the floating toolbar targets your selected nodes.

Select the last node in the chain you want to evaluate, then choose Run Graph. Studio also considers the upstream dependencies needed to supply its inputs. It does not automatically run every other branch in the project. With nothing selected, Run Graph asks you to select nodes. Select all only when you intend to target all those nodes.

Global selection runs skip session nodes and nodes that already have active runs. Session nodes have their own start and stop behavior; start one with its own control. Ordinary node run controls also avoid submitting another run while that node is queued or running.

Before running, check required input values, referenced files, connected services, and any paid operations in the upstream chain. A correctly connected pin is not a guarantee that its value or provider is available.

#Running again or reusing a result

The target tells Studio which result you want. It does not mean that every dependency must execute again.

The concept example targets an unlocked Small preview node, which resizes an image to 0.5 scale. Cached source reuses the unchanged upstream Image source result, illustrated by a blue kite; Small preview runs again. Changed source evaluates the changed source first, illustrated by a coral kite, then runs Small preview with that new image. Locked source reuses the source's frozen blue-kite result, then runs Small preview. This locked example already has a cached output. Other branches are not automatically targets. One node executes when the source is cached or locked; two execute when the source has changed. The artwork and simplified operation labels explain the concept; switching these choices runs no graph and incurs no charge.

A node that needs updating is dirty. A clean upstream node can supply its cached output. Explicitly targeting an unlocked node asks it to run again, even when it is clean. For generative work, a new execution can produce a different result and another charge.

A locked node reuses its cached output even when targeted. If it has no cached output, the run reports an error: unlock it and run it once before locking it. Locking is not a backup of a referenced file; that file still needs to be available.

Auto evaluation can run an eligible, unlocked node after its inputs change. Eligibility includes previously evaluated nodes, and some first-time nodes with media inputs. Do not assume every edit runs every node. Use manual evaluation (turn Auto evaluate off) when you want to decide when another execution starts; Nodes explains the node controls.

#Read progress and results

Submitting work and finishing it are different moments. A node can wait in a queue before its execution starts. Its run control shows a waiting indicator while queued and a progress indicator while evaluating. Some operations also report percentages or log messages; the amount of detail depends on the node.

Independent ready nodes can run together. A downstream node waits for the upstream values it needs. Moving a node on the canvas does not change this order.

Inspect the node's output or preview when it finishes. An old preview can remain useful while you work, but it is not proof that your latest attempt succeeded. Check the node's current state and error message as well.

#Stop work in progress

While work is active, the toolbar's run control becomes Stop Graph. Stopping affects active graph work in that Studio session, rather than only the nodes currently selected.

In In progress, the example has a completed step, a current step running, and a later step waiting. After Stop shows the current step finishing before the later step is stopped. After Force Stop shows the current step interrupted and the later step stopped. The earlier completed result remains in this example. These are conceptual outcomes after the request takes effect, not live progress or an instant-stop guarantee. In a graph with parallel work, normal stopping waits for the current wave of nodes. Stopping cannot undo a request already sent to an external service, and charges already incurred can remain.

Choose Stop Graph to request a normal stop. Already running nodes in the current wave can finish; later work is not started. The stopping popover offers Force Stop when you need to interrupt the running processes instead. Check the resulting state before running again.

Neither option guarantees that an external provider recalls work it has already accepted. Stopping does not by itself mean a refund, a complete result, or a rollback of actions a node already performed.

#If a run fails

Read the failing node's message before retrying. Check the input values and file availability first, then provider sign-in or keys, local-service connections, permissions, and credits where relevant. Fix the specific cause and run the intended target again.

A failed downstream step does not mean that every earlier step failed. Earlier results may already exist, and paid work may already have happened. Reusing a suitable upstream result can avoid repeating expensive generation. See Troubleshooting for the next checks.

#Check cost and keep the result

Where available, a node's footer shows credits used by the last run. That is a record of past execution, not a quote for the next attempt. A missing price or last-run value does not establish that an operation is free. Review Credits and costs before repeating paid work.

Keep the editable workflow and its media deliberately. Save the project when you want to reopen that work, and use the relevant download or file-library action when you need the output itself. A downloaded image does not contain the graph that made it. Saving and recovery and Files and results explain autosaves, saved results, and retained files.