|Docs

Make a UGC ad from product photos with an AI actor

You get a vertical UGC ad, also called a creator-style ad or AI spokesperson ad, where an AI actor talks to camera about your product. Your real product photos and clips sit between the talking shots, so the product on screen is the one you sell. The ad has a hook line, captions and a short call to action.

Use invideo for this when you want the actor, the voice, your product shots and the captions on one editable timeline: you swap the hook, the voice or a product shot and export again, without a new shoot.

Making ads for your own product is ordinary marketing work, and your agent does it. The agent never switches an ad on or spends ad budget. You upload the ad and run it in the ad platform.

Quickest path

Set up your agent first with invideo for AI agents. Then give it this prompt.

Make a 20 second vertical UGC ad for my serum. Use the actor photo./actor.jpg and my product photos in ./product. Write a script withthe hook "I stopped buying three products for this one", make thevoice, animate the actor, cut to my real product shots, add captions.Tell me the price before you generate anything.

In a chat app, you drag the actor photo and product photos into the project in invideo. In a terminal, you name the folder and the agent uploads it.

What you need

The actor and the product come from you. The agent does not invent either one.

  • An actor photo you have the rights to. A front-facing portrait of a person who agreed to appear in AI ads, or a face made with an image model.
  • Product photos and clips. The real product, sharp, from a few angles. A short phone clip of the product in a hand is good.
  • The claim and the hook. What the product does, in your words, and two or three hook lines to test.
  • The shape. 9:16 for TikTok, Reels and Shorts. 4:5 or 1:1 for feeds.
  • The voice. A voice from the library, or your own voice clip.

Steps

The agent works on a branch and tells you what changed at each step.

  1. Set up the project and bring the media in

    The agent makes a project with new_project, a 1080 by 1920 timeline with new_timeline, and a branch with timeline_history. It uploads the actor photo and product files with start_upload and add_media. Expected result: a sequence URL per file.

  2. Write the script and make the voice

    The agent writes a short script: the hook, one or two lines on the product, the call to action. It shows you the script first. It runs find_models with the job in words, picks one, and reads its inputs with get_model. A good pick is elevenlabs_text_to_speech.elevenlabs_tts_v4. It picks a voice with search_library in elevenlabs_text_to_speech_voices, because a voice id works only with its own model. It calls generate with dry_run: true and tells you the price. Then it calls generate with max_credits set to that price and an idempotency_key. Expected result: a sequence URL with the voiceover, which the agent reads with read.

  3. Animate the actor

    The agent runs find_models with the job in words, picks one, and reads its inputs with get_model. A good pick is heygen_photo_avatar, and kling_avatar.std takes the same image and audio. It calls generate with the actor photo as image, the voice as audio, aspect_ratio 9:16 and dry_run: true, and tells you the price. Then it calls generate with max_credits set to that price and an idempotency_key. Expected result: a thread URL and, when the take lands, a sequence URL with a talking video whose length follows the voice. The agent reads it with read and never calls generate again to check.

  4. Keep the product real

    The agent cuts from the actor to your real product photos and clips with place_media. A product shot stays your photo or clip, so the label, color and shape are exact. Expected result: talking shots and real product shots, in script order.

  5. Check the product by eye

    A generated shot can redraw the label or the shape. The agent runs capture on every generated shot that shows the product, and compares each frame with your photo. It replaces any shot where the product changed with your real photo. Expected result: the product looks the same in every frame.

  6. Add the hook, captions and the disclosure

    The agent adds the hook as add_text in the first seconds. It prices audio.transcript on the voice's sequence URL with dry_run, starts it, and calls again after retry_after_s for the words. Then it sends add_captions. It adds a line such as "AI-generated actor" with add_text where the platform asks for one. Expected result: a hook at the top, captions that follow the voice, and the disclosure on screen.

  7. Check and export

    The agent runs capture at the hook, at each cut and at the end, and check_timeline. It checks that captions and the hook sit clear of the app buttons. Then it runs export with format video. Expected result: an export card with status completed.

The same job from a terminal:

invideo project new --title "Serum UGC ad"invideo agent register <project-url> --name <name>export INVIDEO_AS=<name>invideo edit new <project-url> --name "UGC 20s" --width 1080 --height 1920invideo asset upload <edit-url> ./actor.jpg ./product/*invideo library search <project-url> \  --library elevenlabs_text_to_speech_voices --query "warm young woman"invideo gen models <project-url> --query "voiceover from a script"invideo gen model <project-url> \  --model elevenlabs_text_to_speech.elevenlabs_tts_v4invideo gen run <page-url> \  --model elevenlabs_text_to_speech.elevenlabs_tts_v4 --dry-run \  -i prompt=@script.txt -i voice_id=<voice-id>invideo gen run <page-url> \  --model elevenlabs_text_to_speech.elevenlabs_tts_v4 \  --max-credits <price> --idempotency-key serum-voice-1 \  -i prompt=@script.txt -i voice_id=<voice-id>invideo gen models <project-url> --query "talking actor from a photo"invideo gen model <project-url> --model heygen_photo_avatarinvideo gen run <page-url> --model heygen_photo_avatar --dry-run \  -i image=<actor-sequence-url> -i audio=<voice-sequence-url> \  -i aspect_ratio=9:16invideo gen run <page-url> --model heygen_photo_avatar \  --max-credits <price> --idempotency-key serum-ugc-actor-take-1 \  -i image=<actor-sequence-url> -i audio=<voice-sequence-url> \  -i aspect_ratio=9:16invideo status <thread-url> --waitinvideo media audio.transcript <voice-sequence-url> --dry-runinvideo media audio.transcript <voice-sequence-url> --max-credits <price>invideo edit show <edit-url> --jsoninvideo edit change <edit-url> --timeline T1 --token <state_token> \  --operations @ops.jsoninvideo edit capture <edit-url> --at 0.5,4,9,15,19 --out ./framesinvideo edit export <edit-url> --format video --quality 1080 --jsoninvideo status <card-url> --waitinvideo notebook show <card-url> --out ./out

Where the values come from: <project-url> from invideo project new, and <edit-url> from invideo edit new. <name> is the name you choose in invideo agent register. <actor-sequence-url> is the url that invideo asset upload returns for the actor photo. <page-url> is a notebook page from invideo notebook show <project-url>. <voice-id> comes from the library search. <voice-sequence-url> is the sequence URL the voice run returns when it lands. Each <price> is its dry run's estimated_credits. <thread-url> is the url the paid actor run returns. <state_token> comes from invideo edit show. ops.json is a bare list of edit ops. <card-url> is .sequences[0] of what the export returns. invideo status does not download; invideo notebook show saves the file as export.mp4.

The key MCP call makes the talking actor. It is a dry run, so it only checks the inputs and returns the price. <page-url> comes from read on the project URL, and the two sequence URLs come from add_media and the voice generation.

{  "url": "<page-url>",  "model": "heygen_photo_avatar",  "dry_run": true,  "inputs": {    "image": "<actor-sequence-url>",    "audio": "<voice-sequence-url>",    "aspect_ratio": "9:16",    "expressiveness": "medium"  }}

The agent then sends the same call with dry_run false, max_credits set to the price and idempotency_key serum-ugc-actor-take-1.

To change the voice later without a new actor take, the agent calls generate with sync_lipsync_v3.video, the actor video as video and the new voice as audio: a dry run first, then max_credits and an idempotency_key. It sets sync_mode to remap when the new voice is a different length.

Cost and time

Edits and exports are free. Generation and transcripts cost credits. Paid stock costs a small credit charge, or one of your workspace's iStock licenses. Your agent runs a dry run first and tells you the price before it spends anything.

  • Free: uploads, cuts, your product shots, text, captions you place, the checks and the export.
  • Costs credits: the voice, the actor model, any reference-to-video shot, the lip re-sync and the transcript. Actor models charge by the second of audio.

One ad takes under an hour, most of it on the script and the takes you compare. See Credits and what they cost for prices.

If something goes wrong

Most problems are a product that changed in a generated shot, or a mouth out of sync.

SymptomFix
get_model or generate refuses the model as not foundModels retire. Run find_models with the job in words, use the best match, and tell the person which model you used.
Exit code 6, or status is heldA safety review, not a refusal. Wait, then run invideo status <url> --wait again. Do not resend. If it ends failed, change the input the review flagged.
The label or shape changed in a generated shotUse your real photo or clip for that shot. Generated shots are for the actor, not the pack.
You want the actor holding the productGenerate it with seedance_2_5.r2v and your product photos in reference_images: find_models, get_model, a dry run, then generate with max_credits and an idempotency_key. Check the label by eye.
The mouth drifts from the wordsRun sync_lipsync_v3.video on the take with the same voice.
The actor model refuses the photoUse a sharp, front-facing photo of one face. Read the refusal message.
Status needs_confirmationOnly you can confirm it, in invideo.
The download folder is emptyinvideo status does not download. Run invideo notebook show <card-url> --out ./out.
Avoid: the agent resends generate to check on a takeInstead: it reads the thread URL, or calls status.
Avoid: a face of a real person without their consentInstead: use a person who agreed, or a made-up face.

Prompts to try

Each prompt is a real variation of the job.

Make three versions of my UGC ad, each with a different hook fromhooks.txt. Same actor, same product shots. One branch per hook.
Make a 15 second unboxing-style ad: my clip ./unbox.mp4 first, then theactor ./actor.jpg says why they switched. Bold white captions, 9:16.
Swap the voice on my finished UGC ad to a younger male voice, keep theactor and re-sync the lips. Tell me the price first.

When to use something else

For an ad with a real creator, book one on a creator marketplace and bring the clip in with add_media. invideo then cuts, captions and exports it.

The rules for labeling AI actors differ by platform and by country. invideo adds the words you choose, but it does not check an ad against a platform's policy. Read Label AI content so your ad runs on TikTok, Meta and Google before you run the ad.