Pro10 min

Automating Generation End to End

At pro scale you are not clicking generate, you are running a system. A row lands in a sheet, an image is generated, branded, and posted, all without you. This lesson builds that pipeline with n8n as the orchestrator and a vision model to quality check the output.

Step 1: Map the flow

Sketch the pipeline before building. A typical content pipeline: trigger (new row in a sheet or a schedule), generate the image via Flux API, run it past a vision model for a quick quality gate, composite the brand overlay, then publish or save to a folder for review.

  1. Trigger: new row in Google Sheets with a prompt
  2. Generate: call Flux on Fal with that prompt
  3. Gate: ask a vision model if the image matches the brief
  4. Brand: overlay logo and text via a template
  5. Publish: post to the channel or drop in a review folder

Step 2: Add an AI quality gate

The trick that makes automation trustworthy is a vision model checking the work. Send the generated image to Claude Opus 4.8 or Gemini 2.5 Pro and ask a yes or no question against your brief. Only images that pass move forward; the rest get regenerated or flagged.

n8n-quality-gate.json
{
  "node": "AI Vision Gate",
  "model": "claude-opus-4-8",
  "input": {
    "image_url": "{{ $json.flux_image_url }}",
    "question": "Does this image show a single red product on a clean white background with no text? Answer only yes or no."
  },
  "route": { "yes": "Brand & Publish", "no": "Regenerate" }
}
n8n — content pipeline
[Sheets trigger] -> [Flux generate] -> [Vision gate]
| yes -> [Brand overlay] -> [Publish]
| no -> [Regenerate]
A linear flow with one branch: the vision gate sends failures back to regenerate and passes to the brand step.

Step 3: Add a human checkpoint

Full autopilot is risky for anything client facing. Insert a manual approval step: the pipeline drops finished candidates into a Slack channel or a review folder, and only your thumbs up triggers the actual publish. Automate the toil, keep the judgment.

Log everything
Store the prompt, seed, model, cost, and final URL for every run in a database or sheet. When a client asks to redo image 47, you can reproduce it exactly.

Result

You have a pipeline that takes prompts from a sheet, generates and quality checks images automatically, brands them, and parks them for one click human approval before publishing.

Hands-on tasks