The generator bar

Select a node that can generate and a generator bar appears below it. This is the console for one run: write the prompt, pick the model, tune parameters, read the estimate, hit send.

What the generator bar contains

The bar has three layers, top to bottom:

• The top row: the @ button, the add-image button, and thumbnails of the connected references (IMG 1, IMG 2 …); video nodes also have an add-video button;
• The prompt box in the middle: describe the image / video / voice you want; type @ to reference connected nodes, elements or items from your personal library (see References and @ mentions);
• On the left of the bottom row, the model select: models differ in capability and price; you always see the model's display name; with none picked, the default model for that task type is used;
• Next to the model, the parameter summary (such as "Image Generation · 1K" or "16:9 · 1080P · 5s"): click it to set the generate method, aspect ratio, resolution, quality, duration, output format and so on; only parameters the model truly supports are shown. Camera, Presets and similar buttons follow it;
• At the far right of the bottom row, the credit estimate and the send button: the estimate refreshes live as you change things, and send submits the run.

Estimate and billing

Every time you change the prompt, model or a parameter, the bar re-quotes the server and refreshes the estimate, so the price you see and the basis you are actually charged on are identical.

• The estimate shows as one of "— / N credits / up to N credits": models billed per token, per minute or per 5 seconds show "up to", with the real amount settled when the task finishes;
• When a model is unpriced or misconfigured, the estimate slot shows the matching notice rather than a wrong number.

Submit and results

After you hit send, the node goes Queued, then Running, and the result lands back on the node when it finishes.

• Cancel: only while the model service has not started on the request does the node offer Cancel & refund, and cancelling returns all the credits; once the request has gone to the model service it can no longer be cancelled and runs to the end;
• On failure the system decides by its rules whether to refund — only a "Refunded N credits" notice on the node means a refund actually happened; a task settled on actual usage shows "Charged N credits, returned M credits", which is the unused part of the hold, not a refund;
• Double-clicking send never double-charges; there is concurrency dedupe protection.