Blog post turned into 9:16 scene images and a thumbnail

Blog post turned into 9:16 scene images and a thumbnail

An n8n workflow that takes one blog post from Airtable and has an OpenAI model write a four-scene script for a short vertical
video. Leonardo.ai then draws an image for every scene and one for the thumbnail, and the image links go into an Assets table.
The video itself is not made here. You get the script and the pictures to build it from.

What is inside

• workflow.json: the workflow as n8n exports it, 39 nodes, twelve of them sticky notes that label the stages on the canvas.
  It holds no credential, no Airtable id and no saved sample data.
• README.md: these notes.

What we changed from the original copy

• Every credential link is gone: Airtable, OpenAI and the Leonardo.ai Custom Auth credential. Select your own.
• The Airtable base id and the ids of the three tables the workflow names are now placeholders: <YOUR_AIRTABLE_BASE_ID>,
  <YOUR_BRAND_TABLE_ID>, <YOUR_SEO_KEYWORDS_TABLE_ID> and <YOUR_ASSETS_TABLE_ID>.
• Get Content pointed at the same table as the keyword search, but it reads a post by the record id the keyword row links to.
  It now has its own placeholder, <YOUR_CONTENT_TABLE_ID>. Put the id of the table that holds your posts there.
• The saved sample data from the author's test runs is cleared, and the export's tags, instance id and version id are gone.
• One sticky note was an author's promotion. It is now a short title note. The two notes beside the model ids called the
  Leonardo.ai model the latest. They now name it, Phoenix 1.0.

Before you start

• An n8n instance.
• An Airtable base with four tables, described next.
• An OpenAI API key.
• A Leonardo.ai API key with API credits. Leonardo sells API credits apart from its website plans.

The four tables, with the fields the workflow reads and writes:

• Brand guidelines: any rows. Their record ids are read, and no text from them reaches a prompt.
• SEO keywords: Keyword (text) and RelatedContent (a link to a record in the content table).
• Content: Title and Content (text).
• Assets: Asset Name (text), Asset Type (single select with Image), File Size (number) and Asset URL (text).

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Add an Airtable credential. Select it in the five Airtable nodes, and write your base id and the four table ids over the
   placeholders in each of them.
3. Add an OpenAI credential and select it in Script Prep. Its model, gpt-4o-mini, is due to shut down on 14 January 2027, so
   pick another before then.
4. Add a Custom Auth credential for the six Leonardo.ai calls, with this JSON, and select it in Leo - Improve Prompt,
   Leo - Generate Image, Leo - Get imageId, Leo - Improve Prompt1, Leo - Generate Image1 and Leo - Get imageId1:
   {"headers": {"authorization": "Bearer <YOUR_LEONARDO_API_KEY>"}}
5. In Keyword Filter, replace ai automation with a word that appears in the Keyword of the post you want.
6. Press Execute workflow, then look at your Assets table.

How a run goes

1. When clicking ‘Test workflow’ starts the run by hand.
2. Get Brand Guidelines reads every row of the brand table. Set Guidelines keeps each row's id.
3. Get SEO Keywords searches the keyword table once for each brand row and returns Keyword and RelatedContent.
4. Keyword Filter keeps the rows whose Keyword contains ai automation, ignoring case. Remove Duplicates drops repeats of
   the same id, because the search ran once per brand row.
5. Split Out Keywords makes one item for each linked post, and Get Content reads that post by record id. Split Out Content
   and Limit pass on a single post.
6. Script Prep sends the post's title and content to OpenAI, with the Wikipedia tool attached. It asks for a script of four
   scenes that runs under 30 seconds in 9:16. Each scene gets an image prompt, and the thumbnail gets its own. The reply is JSON.
7. Split Out Scenes makes one item for each scene, and Loop Over Items takes them one at a time. Each scene goes through
   Prompt Settings1, which adds a model id, and Leo - Improve Prompt, which asks Leonardo to improve the image prompt. Then
   Leo - Generate Image makes a picture 768 by 1376 pixels with ultra on. Wait 30 Seconds1 waits, Leo - Get imageId
   fetches the finished picture and Add Asset Info1 adds a row to the Assets table. The loop then takes the next scene.
8. Beside the loop, Split Out TN Prompt sends the thumbnail prompt down its own line, through Prompt Settings,
   Leo - Improve Prompt1, Leo - Generate Image1 (alchemy and high resolution on), Wait 30 Seconds, Leo - Get imageId1
   and Add Asset Info.

A run ends with five rows in the Assets table. The four scene rows have an Asset Name that starts with Scene and the scene's
script, and the thumbnail row starts with TN and the post's title. Every row has the image link in Asset URL, type Image
and a file size of 0.

Change the look

• The post: the text in Keyword Filter. Take away Limit's one-item cap to run more posts, and you pay for every one.
• The script: the prompt in Script Prep, which sets the scene count and the length.
• The picture style: the presetStyle in the generate calls, and the image prompt itself. The model id is in the two Prompt
  Settings nodes, and the note beside each says it is Phoenix 1.0.
• The size: width and height in the generate calls.
• The pictures per scene: num_images in the generate calls, which the two optional notes on the canvas point to. Only the
  first picture of each generation is filed in the Assets table.

Good to know

• Models: Script Prep uses gpt-4o-mini. OpenAI's deprecations page, read on 12 November 2026, lists the gpt-4o-mini
  snapshot gpt-4o-mini-2024-07-18 for shutdown on 14 January 2027, with gpt-6-luna as the replacement. Pick another model in
  the node before then.
• Names: the prompt gives the model a layout, not key names. The next steps look for a scenes list with script and
  image_prompt in each scene, and for thumbnail_prompt. If a run stops at a Split Out node, add those names to the prompt.
• The thumbnail's improve call reads the field message.content['Thumbnail Prompt'], while the split step before it is set to
  message.content.thumbnail_prompt. If the thumbnail prompt arrives empty, make the two names the same.
• The brand table does not shape the output. Its rows are fetched, and none of their text reaches a prompt. The additional
  field is set and not used.
• Cost: a run makes one OpenAI call and fifteen Leonardo.ai calls, five of them generations: four scenes and the thumbnail. The
  scenes use ultra and the thumbnail uses alchemy with high resolution, which cost more than a plain generation. Check
  Leonardo's pricing page for the current rate.
• Time: each generation is followed by a fixed wait of half a minute. If a picture is not ready by then, the fetch can return no
  image, and the Airtable step then fails. Raise the wait.
• The API: the workflow calls Leonardo's v1 generations endpoint, which its documentation still lists on 12 November 2026.
• Your data: the post goes to OpenAI, the prompts go to Leonardo.ai, and the generations are set to private on your Leonardo
  account.
• One post per run, on purpose: Limit passes a single item.