Stock photo title, description and keywords from a webhook

Stock photo title, description and keywords from a webhook

An n8n workflow with a webhook. You post it a picture, either as a public address or as base64, and gpt-4o-mini
looks at the picture and writes a title, a description and a list of keywords for a stock marketplace. The reply holds
them as fields and as a ready CSV row.

What is inside

• workflow.json: the workflow as n8n exports it, 7 nodes and no sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

The Authorization header of Vision LLM held the text Bearer YOUR_OPENAI_API_KEY. It now reads
Bearer <YOUR_OPENAI_API_KEY>, so a key pasted over the brackets is easy to spot. The rest is as the original copy had it.

Before you start

• An n8n instance with an address the sender can reach. n8n Cloud works, and so does your own server on the internet.
• An OpenAI API key that can use gpt-4o-mini.
• Pictures to send: public links, or the files themselves as text.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Open Vision LLM. In its Authorization header, replace <YOUR_OPENAI_API_KEY> with your key. The brackets go. Keep
   the word Bearer and the space after it.
3. Save, then open Webhook and copy its test address. It ends in stock-image-meta.
4. Press Listen for test event, then post a picture to that address. This body works:

{ "imageUrl": "https://example.com/photo.jpg" }

1. Press Publish when the reply looks right. Your production address is the same path under /webhook/.

How a run goes

1. Webhook takes a POST on the path stock-image-meta. It waits for a node to answer, so the sender gets the result
   in the same call.
2. Parse Input reads imageUrl, image_url or image from the body. It also accepts imageBase64, and when that
   text has no data: start it adds data:image/jpeg;base64,. With no picture it stops with an error that says so.
3. Build Vision Request writes the request for OpenAI's chat completions API. The system message makes the model a
   stock photo metadata writer. The user message asks for raw JSON with a title under 12 words, a 2 to 3 sentence
   description and 20 to 40 lowercase keywords. It sets gpt-4o-mini, a temperature of 0.3 and 800 tokens.
4. Vision LLM posts that body to https://api.openai.com/v1/chat/completions with a 120 second timeout.
5. Parse Metadata takes the model's text, removes code fences, reads the JSON, and falls back to the first {...}
   block if that fails. A missing title becomes Untitled. A missing description or an empty keyword list stops the run.
6. Build CSV Row doubles any quote marks, joins the keywords with a semicolon and a space, and builds csv_row and a
   csv text that starts with the header "title","description","keywords".
7. Respond to Webhook answers with every incoming item as JSON.

Change the look

• The metadata's rules: the two messages in Build Vision Request. Change the title length, the keyword count or the
  marketplace it names.
• The model: model in the same node. A newer model may refuse max_tokens or temperature, so check that it accepts
  both before you swap it in.
• The answer's length: max_tokens in that node.
• The CSV: Build CSV Row. Change its separator to a comma if your marketplace wants one.
• The address people post to: path in Webhook.

Good to know

• OpenAI's deprecations page, read on 12 November 2026, lists the dated gpt-4o-mini version for shutdown on 14
  January 2027, with gpt-6-luna as the replacement. It does not list the bare name gpt-4o-mini, which this copy
  keeps.
• Respond to Webhook is set to answer with all incoming items, so the reply is a JSON array with one object holding
  title, description, keywords, keyword_count, csv_row and csv. The responseBody text in that node, which
  would add a success flag, is not used with this setting.
• The 20 to 40 keywords are only a request in the prompt. Nothing counts them, and keyword_count reports what came
  back.
• An address in imageUrl must be reachable by OpenAI, so a link behind a login or on a private network fails. A bare
  base64 text is always labelled image/jpeg, whatever the file is.
• The webhook has no authentication. Anyone with its address can start a call that spends your key, so turn on the
  webhook's own authentication or keep the address private.
• One gpt-4o-mini call runs for each picture, and every picture goes to OpenAI.
• The key sits in a header field, not in a credential, so it travels with any export. Take it out before you share the
  workflow.