UGC-style video ad from a product and a creator photo

UGC-style video ad from a product and a creator photo

An n8n workflow that makes a casual, phone-shot video ad. Send a form with two photos, usually your product and a
person, and one line of direction. WaveSpeedAI blends the photos, two agents on gpt-4.1 write an image prompt and the
scenes, Kie.ai renders a still and then Veo clips, and fal.ai joins the clips into one video.

What is inside

• workflow.json: the workflow as n8n exports it, 38 nodes, five of them sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The ten Authorization headers held bare placeholders. They now read <YOUR_DROPBOX_TOKEN>, <YOUR_WAVESPEED_KEY>,
  <YOUR_KIE_KEY> and <YOUR_FAL_KEY>, each in the nodes that call that service.
• Four repairs: the original copy had lost its backslashes.
• Create Image1 and Create Video1 escape the prompt before it goes into the JSON body. Their replace calls had
  become /n/g and '"', so every letter n turned into a line break and the quotes went out bare, which breaks the
  body. They read .replace(/\n/g, '\\n') and .replace(/"/g, '\\"') again.
• The path in Upload a file swaps spaces in a file name for underscores. Its /s+/ is /\s+/ again.
• The examples in the two parsers and in the two agents' system messages show a YAML prompt as a JSON string. Their
  line breaks had become a bare n between words, as in naturallyncharacter. They read \n again.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• An OpenAI API key for gpt-4.1 and for a vision model.
• A Dropbox account: a Dropbox credential in n8n for the upload, and an access token for the two sharing calls.
• A WaveSpeedAI API key, for FLUX Kontext dev's multi-image model.
• A Kie.ai API key with credit, for its 4o Image model and for Veo.
• A fal.ai API key, for its FFmpeg merge.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Add an OpenAI credential and select it in Describe Img1 and GPT1.
3. In Describe Img1, pick a current model that reads images, such as gpt-4.1, in place of chatgpt-4o-latest.
   OpenAI's deprecations page, read on 2 November 2026, lists that one as shut down on 17 February 2026.
4. Add a Dropbox credential and select it in Upload a file. In HTTP Request4 and HTTP Request6, replace
   <YOUR_DROPBOX_TOKEN> in the Authorization header with a Dropbox access token. The brackets go too, so the header
   reads Bearer, a space and the token.
5. In HTTP Request and HTTP Request1, replace <YOUR_WAVESPEED_KEY> the same way.
6. In Create Image1, Get Image1, Create Video1 and Get Video1, replace <YOUR_KIE_KEY>.
7. In Combine Clips1 and Get Final Video1, replace <YOUR_FAL_KEY>. That header reads Key, a space and the key.
8. Under the workflow's Settings, set a timeout, for the reason under Good to know.
9. Save, open the form's test address, attach two photos and write one line. Press Publish when a clip comes back.

How a run goes

1. On form submission takes UGC Instruction and Images. Split Out and Code2 turn each photo into its own
   item, named product_image.
2. Upload a file puts each photo in your Dropbox. HTTP Request4 looks for a shared link, HTTP Request6 makes a
   public one, and Code7 turns the links into raw file addresses, gathered in one list.
3. HTTP Request asks WaveSpeedAI to blend the first two photos into one picture. Wait holds 20 seconds, then
   HTTP Request1 fetches it.
4. Describe Img1 describes the picture in YAML: a product, a character or both.
5. Generate Image writes one image prompt and picks 2:3 or 3:2. Create Image1 asks Kie.ai's 4o Image model for a
   new still from the blended picture. Wait 4 and If 1 check once a minute until it is ready.
6. Generate Video writes the scenes, one for each 8 seconds of the length you asked for, one when you name none.
   Each scene has its prompt with the spoken line, a frame (9:16 unless you ask for 16:9) and a model (veo3_fast
   unless you ask for veo3).
7. Split Out1 sends the scenes to Create Video1, one every 3 seconds, which asks Kie.ai's Veo to animate the still.
   Wait holds a little over three minutes, and Get Video1 reads each clip once.
8. Aggregate1 gathers the clip addresses, and Combine Clips1 sends them to fal.ai's merge. Wait1 and If 3 check
   once a minute until the merged video has an address.
9. Code1 lays out the video's address, name and size with ready HTML and Markdown snippets, and HTTP Request2
   downloads the file into the run.

Change the look

• The still: the system message of Generate Image, which asks for amateur phone-photo realism and exact product text.
• The script: the system message of Generate Video. It writes a casual line under 150 characters when you give none,
  and mentions the brand only in the first scene.
• Length, frame and model: say them in your instruction. The agent reads them from there.
• The blend: the prompt in the body of HTTP Request.
• The waits: Wait 20 seconds, Wait 4 and Wait1 60 seconds, Wait 200 seconds. Raise Wait for veo3,
  which takes longer.

Good to know

• WaveSpeedAI bills the blend, Kie.ai the still and every clip, fal.ai the merge, and OpenAI the description and the
  two agents. A long ad means more clips.
• Send exactly two photos. The blend takes the first two links and nothing else, so one photo leaves an empty address
  and a third is ignored.
• A photo whose file name already has a Dropbox link takes the first output of If, which leads nowhere, and the run
  goes on without it. Rename the photo, or add {{ $now.toMillis() }} to the path in Upload a file.
• A failed image never meets the test in If 1, so the run checks it again every minute for as long as n8n lets it.
  That is why step 8 is there.
• Get Video1 reads each clip once, after 200 seconds, with no status check. A clip still rendering has no address
  and drops out of the merge.
• An access token made in Dropbox's App Console expires after a few hours. For a lasting setup, set both sharing calls'
  Authentication to a Dropbox OAuth2 credential and delete the header.
• OpenAI's deprecations page lists gpt-image-1, the GPT-4o image model in its own API, as shut down on
  23 October 2026. Kie.ai runs its own 4o Image service, so check that Create Image1 still answers before a client
  run, and switch to another of Kie.ai's image models if it does not.
• The same page lists GPT-4.1 (gpt-4.1-2025-04-14) for removal from the API on 14 January 2027, with gpt-6-luna
  as its replacement. GPT1 drives both agents.
• Once published, anyone with the form's address can start a run on your five accounts.
• Your photos become public on Dropbox, and they go to WaveSpeedAI, OpenAI and Kie.ai.
• Your keys sit in ten headers, so take them out before you share the workflow.