Quickstart

This walks through one generation end to end. It uses Text to Speech because it is the fastest app to complete — a job finishes in seconds rather than minutes, so you see the whole loop before committing time to a video render.

The loop is identical in every other app.

Before you start

You need an account and a non-zero credit balance. New workspaces start with a trial balance; check yours in the top-right of the dashboard.

Run your first generation

1. Create a project.

   From the dashboard, select New project. A project is a container for related generations and their assets — everything you make in this walkthrough lands inside it.

   Name it something you will recognise in a week. Projects are the unit you search by later, and Untitled project 4 is not findable.

2. Open an app.

   Go to Audio → Text to Speech. Every app has the same three-part layout: inputs on the left, the output canvas in the middle, and parameters on the right.

3. Fill in the input.

   Paste a couple of sentences of real script. Something with punctuation and a natural rhythm — a delivery model has nothing to work with in test test test, and the output will sound like it.

4. Check the estimate, then generate.

   The credit cost appears next to the Generate button before you run anything. It is not an approximation for most apps — it is the exact charge for the parameters currently set.

   Select Generate.

5. Wait for the job.

   The job enters the queue and the canvas shows its progress. You can leave the page: jobs run server-side and appear in your Library when they finish. Nothing is lost by navigating away.

6. Export.

   When the job completes, use the download control on the output. Choose the format you actually need — re-exporting later is free, but re-generating is not.

Judge the model, not the render

Generate two or three variations before changing your prompt. Most apps produce a meaningfully different result on each run from the same input, so a single output tells you very little about whether the prompt is working.

What just happened

Four things, worth naming because the rest of the docs assume them:

• A job was queued, run and stored. Jobs are durable — closing the tab does not cancel one.
• Credits were deducted at completion, not at submission. A job that fails on our side is not charged.
• The output landed in your Library, scoped to the project.
• The generation kept its parameters, so you can reopen it and re-run with one value changed rather than rebuilding the input.

Next

If the output was not what you wanted, that is normal on a first attempt and it is nearly always a prompting problem rather than a model choice problem. Read the prompting guide for the studio you are working in before you start swapping models.

If you will be generating anything involving a consistent person, product or visual identity, set up a brand kit first. It is the difference between reference-matching once and re-uploading the same files on every generation.