NewHow we catch a seller listing the wrong model, and refund you if one does.
UberLLM

Ad creative workflows

Four recipes that turn one product photo into ad creatives, and the pattern they share. Each recipe is an n8n workflow you import by link and one command of a script with no dependencies. What each one made and cost on a real run is on the ad creative page.

Recipe n8n workflow (import by link) Script command
A person holding the product (9:16 still), a clip from that still, a spoken line ad-ugc-still-clip-voice.json ugc product.jpg
Static variations, one picture per line of a list ad-static-variations.json variations product.jpg
One still and one clip per row of a CSV file or Google Sheet, on a schedule ad-sheet-rows-to-ads.json sheet rows.csv
A studio still of the product alone and a turntable-style clip ad-product-spin-video.json spin product.jpg

The files are in the templates repository: workflows in workflows/, the script and the agent skill in skills/ad-creative/.

Run one

In n8n. Import the link (⋯ → Import from URL…), create one Header Auth credential (Name Authorization, Value Bearer ull_...), edit the Settings node and click Execute workflow. Setup in full: n8n video templates.

As a script (Node 18 or newer):

export UBERLLM_API_KEY=ull_...
curl -O https://raw.githubusercontent.com/PSA-Team-source/uberllm-n8n-templates/main/skills/ad-creative/scripts/ad-creative.mjs
node ad-creative.mjs ugc product.jpg --dry-run   # prices the run; makes nothing, charges nothing
node ad-creative.mjs ugc product.jpg --merge     # --merge joins clip and voice with ffmpeg

From a coding agent (Claude Code and others that read skills):

npx skills add PSA-Team-source/uberllm-n8n-templates --skill ad-creative

The pattern every recipe follows

1. Make the still in the shape you want

curl https://api.uberllm.dev/v1/images/edits \
  -H "Authorization: Bearer $UBERLLM_API_KEY" -H "Content-Type: application/json" \
  -d '{ "model": "google/nano-banana-2-edit",
        "prompt": "A hand holds this exact product in a bright kitchen, phone-camera look",
        "image_url": "https://example.com/product.jpg",
        "aspect_ratio": "9:16", "n": 1 }'

The answer has the picture at data[0].url. The image-to-video models in these recipes make the clip in the shape of its first picture and take no aspect_ratio (a request that sends one is refused before any charge), so the shape is chosen here: 9:16, 4:5 or 1:1.

A picture request answers when the picture is ready. It can time out with a 504, which is not charged: send it again. The recipes try up to three times.

2. Keep the still, and hand it to the video model as a file

The edited picture comes back as a link to /v1/media/artifacts/…, which needs your key and lasts about 15 minutes. Download it at once, then upload it and use the file id as the clip's image_url:

curl -o still.png -H "Authorization: Bearer $UBERLLM_API_KEY" "<data[0].url>"
curl https://api.uberllm.dev/v1/files -H "Authorization: Bearer $UBERLLM_API_KEY" -F [email protected]
# → { "id": "file_…", … }   kept for 24 hours

Do not pass a /v1/media/artifacts/… link as image_url of a video job. A provider cannot open a link that needs your key, and the job fails (at no charge). A file id or a /v1/files/… link works.

3. Start the clip, with a ceiling

curl https://api.uberllm.dev/v1/video/generations \
  -H "Authorization: Bearer $UBERLLM_API_KEY" -H "Content-Type: application/json" \
  -H "Idempotency-Key: ad-0042-clip" \
  -d '{ "model": "kling/kling-o3-standard-image-to-video",
        "models": ["fal-ai/kling-video/o3/standard/image-to-video"],
        "prompt": "The hand lifts the product toward the camera and turns it a little",
        "image_url": "file_…", "duration_seconds": 5, "audio": false,
        "max_cost": 1.5 }'
  • The answer comes at once with the job id, a poll_url and estimated_cost_usd.
  • Idempotency-Key: send the same request again with the same key (after a dropped connection, or an n8n retry) and you get the same job back, not a second one.
  • models: backups, tried in order when the first has no free provider.
  • max_cost: your ceiling in US dollars. A job that would cost more is refused or cancelled at no charge. The bill is what the provider charged, which can be above the price shown at the start when a job moves to another provider, so set this on every job.
  • Add "validate": true to check and price the request without making anything.

4. Wait: poll or take a callback

Poll. Ask poll_url every 5 to 10 seconds. It answers 202 while the job is queued or running, and 200 once it has ended: succeeded, failed, canceled or expired. The last three cost nothing. In the test runs a 5-second clip took between 1.5 and 5.5 minutes, so allow at least 15 minutes before giving up; a job nobody checks is still finished and billed, or refunded. All statuses: Job status.

Callback. Add "callback_url": "https://your-app.example/uberllm/jobs" to the start request and UberLLM sends one POST with the job's JSON when it has ended and its bill is final. In n8n this is a second workflow that begins with a Webhook node. Check the UberLLM-Signature header before you trust a message: Check the signature.

File Link lasts
A picture at /v1/media/artifacts/… about 15 minutes
A video about 3 hours after the job ends
A file you uploaded to /v1/files 24 hours

After that the link answers 404 artifact_expired. The recipes download every file as soon as it exists. To have a clip put straight into your own storage, add output_upload_url (a presigned PUT link) to the start request: Send results to your own storage.

6. Voice, captions and the join

curl https://api.uberllm.dev/v1/audio/speech -o voice.mp3 \
  -H "Authorization: Bearer $UBERLLM_API_KEY" -H "Content-Type: application/json" \
  -d '{ "model": "venice/kokoro-tts", "voice": "af_heart", "response_format": "mp3",
        "input": "Cold water all day, wherever the day takes you." }'

# a timed caption file from that voice track
curl https://api.uberllm.dev/v1/audio/transcriptions -o voice.srt \
  -H "Authorization: Bearer $UBERLLM_API_KEY" \
  -F [email protected] -F model=venice/whisper-large-v3 -F response_format=srt

# UberLLM does not join files: clip + voice is one ffmpeg command
ffmpeg -i clip.mp4 -i voice.mp3 -map 0:v:0 -map 1:a:0 -c:v copy -c:a aac -af apad -shortest ad.mp4

The speech answer is the audio itself; its cost is in the X-UberLLM-Cost-USD header.

Many ads from a sheet

The sheet recipe reads a CSV with the columns name, image_url, scene, motion and aspect_ratio (a sample). A Google Sheet works once it is published as CSV (File → Share → Publish to web → CSV).

  • It makes the rows it has not made before, at most 5 per run unless you change it.
  • A row that fails does not stop the others and is tried again on the next run.
  • In n8n the memory of finished rows is saved when the workflow is active and runs from its schedule. A manual test run does not save it. The script keeps it in done.json in the output folder, so it works from cron.

What these recipes do not do

  • Join clips, or put a voice track or captions on a clip (that is the ffmpeg step).
  • Talking avatars or lip-sync.
  • Keep your files: download them.
  • Check the result. A model can change a label, a colour or a shape, and it draws the sides your photo does not show. Look at every picture and clip before it runs.

Before you run ads

  • People in these pictures are AI-generated. Label them where the platform or your local law asks for it, and never present one as a real customer, a testimonial or an expert.
  • Keep health claims and before/after pictures out of prompts and voice lines, and follow the ad platform's rules for your product's category.
  • Each model has a maker whose terms say how its output may be used. Read them for the models you use, and use only product photos you have the rights to.

Sources for these rules are linked on the ad creative page.