AI UGC ads:
product shots, a presenter and a voiceover from one API key
Make UGC-style video ads from code: product photos from the images API, a presenter who talks in video shots or a swap, and a voiceover from text to speech.
Last updated 2026-11-05
Black Friday deposit bonus
Every qualifying top-up earns free credit, on top of any other offer.
Ends November 30
A UGC-style ad is a person holding your product, talking to the camera like a friend, cut into a few short vertical clips. You can make the parts of one from code, with one prepaid key: the product shots with the image generation API, the presenter with video shots (a saved character who speaks your script in their own voice) or with a character swap over footage you filmed, and an optional voiceover with text to speech. This page says which API makes which part, and what is still yours to do: the script, the checks, the edit and the disclosure.
Hand it to your AI agent
Copy this prompt into Claude Code, Cursor, Codex or any coding agent. It tells the agent where the docs are and how to use your API key.
Build a pipeline that turns product photos and an ad script into vertical UGC-style video ads with the Lovely Router images, video shots, joins and speech APIs. - Docs (read first): https://lovelyrouter.com/docs/models/images.md and https://lovelyrouter.com/docs/api/video-shots.md and https://lovelyrouter.com/docs/api/audio.md - Images: POST https://lovelyrouter.com/v1/images/edits with the product photos as references. Shots: POST https://lovelyrouter.com/v1/videos/shots (`consent: true`; read the model licence in the docs and tell me if my use or region might not be allowed). Voiceover: POST https://lovelyrouter.com/v1/audio/speech. Join with POST https://lovelyrouter.com/v1/videos/joins. - Auth: read my API key from the LOVELY_API_KEY environment variable and send it as `Authorization: Bearer $LOVELY_API_KEY`. Never print or commit the key. If it is not set, ask me to create one at https://lovelyrouter.com/app/keys. - Billing is prepaid: every call is charged to my balance, so test with a small request before running anything large. - Every docs page is plain Markdown; the full index is https://lovelyrouter.com/llms.txt. My task: <the product photos, the presenter, the script and the languages>
How it works
Shoot the product without a studio
Send a photo of your product to
POST /v1/images/editsas a reference image and describe the scene: on a bathroom shelf, in a gym bag, in a hand by a window. Ask for up to 10 versions in one request withn, at one flat price per image.Cast a presenter
Draw a turnaround sheet of your presenter with the images API and save it as a character, free. Give the character a voice: design one from a description with
qwen3-ttsand save it, or upload a voice that is yours or that you have permission to use.Film the lines as shots
Each clip is a shot:
POST /v1/videos/shotswith a prompt that tags the presenter (@Jess) and puts the line in double quotes. The shot comes back with picture, sound and the line spoken in the character's voice. To open on the product, send a frame as a scene and say the shot starts on it.Join the cut
POST /v1/videos/joinsputs up to 50 finished shots of the same size together in order, each with its own sound, for free. What you add after that (captions, music, a logo end card) happens in your own editor.
The requests
Product shots first, with the OpenAI SDK pointed at https://lovelyrouter.com/v1:
import base64
import os
from openai import OpenAI
client = OpenAI(base_url="https://lovelyrouter.com/v1", api_key=os.environ["LOVELY_API_KEY"])
result = client.images.edit(
model="gpt-image-2.5-flare",
image=[open("serum.png", "rb")],
prompt="The serum bottle from the image on a sunlit bathroom shelf next to a folded towel, phone-camera look, vertical",
size="1024x1536",
n=4,
)
for i, image in enumerate(result.data):
with open(f"serum-{i + 1}.png", "wb") as f:
f.write(base64.b64decode(image.b64_json))Then a shot: the presenter is a saved character, and the opening frame (for example one of your product shots edited with the presenter's sheet as a second reference image) is sent as a scene.
curl "https://lovelyrouter.com/v1/videos/shots?wait=300" \
-H "Authorization: Bearer $LOVELY_API_KEY" \
-H "content-type: application/json" \
-d '{
"model": "minimax-h3-video",
"consent": true,
"size": "9:16",
"seconds": 10,
"prompt": "The shot opens exactly on @[Opening frame] as its first frame. @Jess picks up the bottle, holds it to the camera and says: \"Okay, this is the one I keep telling you about.\"",
"characters": ["chr_3b9e1f0c7a2d4e58b6c1a9f2"],
"scenes": [{"name": "Opening frame", "image": "https://example.com/jess-with-serum.png"}]
}'Make the other lines the same way, then join them with POST /v1/videos/joins and {"shots": ["shot_…", "shot_…", "shot_…"]}. The video shots reference covers polling, the tags and the errors.
Which API makes which part
| Part of the ad | Made by | Notes |
|---|---|---|
| Product shots and the opening frames | Images API: gpt-image-2, gpt-image-2.5-flare, gpt-image-2.5-sunburst | Up to 16 reference images per edit, n up to 10, up to 3840×2160 |
| The presenter's look | A turnaround sheet from the images API, saved with POST /v1/characters (sheet_image_job) | Free to save, up to 100 per account |
| The presenter's voice | qwen3-tts designed voice saved with POST /v1/audio/voices, or an uploaded sample with consent | Linked to the character as voice_id |
| Talking clips | Video shots with minimax-h3-video | 5–15 s each, 9:16, 16:9 or 1:1, sound and dialogue included |
| A presenter in your own footage | Character replacement with mode: "face" | Film once, swap the presenter's head per market; hands, clothes and product stay as filmed |
| A separate voiceover | Text to speech | An audio file; you mix it over the video yourself |
| The final cut | POST /v1/videos/joins, free | Joins shots in order; no captions, music or overlays |
What you still do
The shot speaks the words you put in quotes, word for word. What your ad promises is your responsibility.
A generated image or shot can change small things: the label text, the cap, the colour of the liquid. Look at every frame against the real product before you publish, and regenerate the ones that drifted.
Joins give you one mp4 with the shots' own sound. Captions, music, a voiceover mix, your logo and the end card are done in your editor or with your own ffmpeg step.
Say the presenter is AI-generated wherever the ad runs, follow the ad platform's rules, and never present a generated presenter as a real customer or a real review.
The video models are under the MiniMax H3 terms, which every request accepts with consent: true. They don't allow use in the EU, the UK, South Korea or the US.
Recipes
Loop over your catalog and send each product photo with three scene prompts and n: 4, as image jobs so nothing holds a connection open.
Save the presenter once and reuse the character_id: same face, same clothes, same voice across a whole campaign.
Send language (en, zh, ja, …) on each shot, with the line translated in quotes.
Film the product demo yourself, then swap the presenter's head with character replacement for each market, keeping your hands and the product exactly as filmed.
Render several 5-second opening shots with different first lines, join each one to the same body, and test them as separate ads.
What you pay
| Part | Price |
|---|---|
| Product shots, sheets and frames | $0.04 per image, any size, generations and edits alike |
| Video shots | $0.13 per rendered second with one picture, plus $0.06 per second for each extra picture |
| Joins, saved characters, saved voices | Free |
| Voice design and voiceovers | $0.036 per minute of audio |
| Character replacement | $0.2 per second swapped |
A 30-second ad made of three 10-second shots, each with the presenter and an opening frame (two pictures), renders 10.13 seconds per shot and costs about $5.77 for the video, plus a few images at $0.04 each. When a shot has two or more characters, a short chat call names everyone in its prompt and is billed as an ordinary chat request. Shots are reserved from your prepaid balance when you submit them and charged when they complete; failed shots cost nothing. Every price is on pricing.
Frequently asked questions
Is the whole ad generated in one call?+
No. Each part is its own request: images for the product shots, a shot for each clip, a join for the cut. That keeps every step checkable, and you only regenerate the part that missed.
Does the presenter actually talk?+
Yes. Words in double quotes in a shot's prompt are spoken by the character tagged before them, in that character's saved voice when it has one, with the sound made by the video model.
Will my product look exactly like the real one?+
Close, but check it. Reference images keep the shape and colours well, but small text and fine details can change. Compare every frame to the real product and regenerate the ones that drifted.
Can I use my own footage instead of a generated presenter?+
Yes. Film the ad yourself and swap the presenter with character replacement, as a whole person or only the head. The motion, the camera, the cuts and the sound stay.
Can I add music and captions through the API?+
No. Joins put shots together with their own sound. Music, captions, overlays and a separate voiceover are added in your editor.
Where can I run these ads?+
Follow the ad platform's rules and the model's terms. The MiniMax H3 terms don't allow use in the EU, the UK, South Korea or the US, and every generated presenter should be disclosed as AI.
Can an AI agent build this pipeline?+
Yes. Copy the prompt on this page into Claude Code, Cursor or Codex: it reads the Markdown docs and writes the image, shot and join calls with your key.
For still product photography only, see AI product photos; to try a shot first, open the video playground, or create an account and get a key.