Video narration script and voice-over from sampled frames

Video narration script and voice-over from sampled frames

An n8n workflow that narrates a video. It downloads a video, picks up to 90 frames spread evenly over it and sends them in
batches of 15 to an OpenAI vision model, which writes the narration script one batch after another. The finished script
goes to OpenAI text-to-speech, and the mp3 voice-over is uploaded to Google Drive.

What is inside

• workflow.json: the workflow as n8n exports it, 14 working nodes and seven sticky notes. It holds no credential.
• README.md: these notes.
• The sample video is by Coverr-Free-Footage on Pixabay.

What we changed from the original copy

• Capture Frames is a Python node. The original copy had lost its indentation, with every line at the same level, so it
  could not run. It also read the video with _input.item.binary.data.data, which n8n's Python runner does not offer. The
  script is indented again and reads _item["binary"]["data"]["data"]. Its import of NumPy, which it never used, is gone.
• Python in n8n has changed. The Code node page, read on 12 November 2026, says the old Python option, Pyodide, is removed
  from n8n 2.0, and that n8n Cloud does not allow Python imports. This step needs OpenCV, so it needs your own server and
  a Python runner that has OpenCV. See Before you start.
• OpenAI Chat Model held gpt-4o-2024-08-06. OpenAI's deprecations page, read on 12 November 2026, lists it for shutdown
  on 14 January 2027, with gpt-6-luna as the replacement. The node is now version 1.2 and the model is gpt-6-luna. Its
  model page, read the same day, lists image input and chat requests.
• Stay Within Service Limits had no values set, so it used n8n's defaults, which are one hour. It now waits 30 seconds.
• The prompt asked for the style of a named broadcaster. It now asks for the style of a nature documentary narrator.
• Upload to GDrive held its author's folder ID. It now holds YOUR_FOLDER_ID.
• The seven sticky notes are rewritten and hold no outside links.
• Every credential reference is removed, so you pick your own.

Before you start

• Your own n8n server. n8n Cloud cannot run the frame step, because it does not allow Python imports.
• A Python task runner that has OpenCV. Set it up in step 2 below.
• An OpenAI API key that can use gpt-6-luna and OpenAI's speech endpoint.
• A Google account, a Drive folder and a Google Drive OAuth2 credential.
• GraphicsMagick for the Edit Image node. n8n's own Docker image includes it, and a server you set up yourself needs it
  added.
• Enough memory. The video and its frames are held in memory, and a long video can stop your n8n.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Give the Python runner OpenCV. n8n's task runner page, read on the same day, shows how to build a runner image with
   extra Python packages and how to allow them. Install opencv-python-headless in that image, set
   N8N_RUNNERS_EXTERNAL_ALLOW to cv2 and set N8N_RUNNERS_STDLIB_ALLOW to base64. The runner's source, read on the
   same day, blocks open by default, and this script uses it to save the video before reading it. Set
   N8N_RUNNERS_BUILTINS_DENY to the default list without open.
3. Add an OpenAI credential and select it in OpenAI Chat Model and Use Text-to-Speech.
4. Add a Google Drive OAuth2 credential and select it in Upload to GDrive. Replace YOUR_FOLDER_ID with the ID of your
   folder. It is the part of the folder's address after /folders/.
5. To use your own video, change the address in Download Video. Pick a short one in a format OpenCV can read.
6. Run Download Video and then Capture Frames by themselves first. When they return frames, press Test workflow.

How a run goes

1. Download Video downloads the sample mp4, about 3 MB.
2. Capture Frames saves the video as a temporary file, steps through it at an even spacing and returns each frame as
   base64 text of a jpeg, in one list.
3. Split Out Frames makes one item for each frame.
4. For Every 15 Frames takes 15 frames at a time. Its done line goes to Combine Script, and its loop line goes to
   Convert to Binary.
5. Convert to Binary turns each text into an image file. Resize Frame fits it into 768 by 768 pixels with its
   proportions kept, as a jpeg. Aggregate Frames puts the 15 images on one item.
6. Generate Narration Script attaches the 15 images and asks OpenAI Chat Model for a short voice-over script. From the
   second round on, the prompt also holds the script written so far, so the story continues.
7. Stay Within Service Limits waits 30 seconds, and the loop takes the next 15 frames.
8. Combine Script gathers the parts. Use Text-to-Speech joins them with line breaks and makes an mp3. Upload to GDrive
   saves it in your folder as narrating-video-using-vision-ai- and the time, ending in .mp3.

Change the look

• The narration style: the prompt in Generate Narration Script.
• How many frames: max_frames in Capture Frames.
• The batch size: For Every 15 Frames. The chain has 15 image slots, so a bigger batch needs more slots there.
• The frame size: the width and height in Resize Frame.
• The voice: the model and voice in Use Text-to-Speech.
• The pause: Stay Within Service Limits.
• The video: the address in Download Video.

Good to know

• Capture Frames is the hard step. The OpenCV calls are the original's, and only the indentation and the way the video is
  read changed, so run it by itself on the sample video before you rely on it.
• gpt-6-luna is the replacement that OpenAI's deprecations page names for the dated gpt-4o version in the original.
• Use Text-to-Speech sets no model, so n8n uses its default, tts-1. The same deprecations page lists tts-1, tts-1-hd and
  gpt-4o-mini-tts for shutdown on 6 January 2027, with gpt-realtime-2.1-mini as the replacement. The text to speech
  guide, read on the same day, says that model is not supported by the speech endpoint this
  node calls. The node keeps working until the shutdown, so plan to replace this step before then.
• OpenAI's usage policies, read on the same day, require you to tell listeners the voice is AI-generated.
• n8n's node says the text for speech can hold at most 4096 characters. Six short parts usually fit, but a longer script
  needs splitting.
• OpenAI bills six chat requests with 15 images each, and one speech request, for a 90-frame run.
• The waits add three minutes to a run. Shorten or remove Stay Within Service Limits if your rate limits are high.
• The frames, the script and the narration prompt go to OpenAI, and the finished mp3 goes to Google Drive.