Use cases

Building a UGC ad generator with AI avatars

How to build a UGC ad generator with AI avatars: creator clips, hook and CTA variant matrices, captions, costs per variant and disclosure, for developers.

PitchTwin team, , 6 min read

Performance marketers want dozens of ad variants a week, and UGC-style creative is what they keep asking for. Real creators are slow to book and expensive to reshoot for every new hook, while generic AI faces look synthetic in a feed where authenticity is the whole point. A UGC ad generator sits in the middle: a small roster of real creators who record once, and software that turns scripts into many variants.

This guide walks through the product and technical design: the creator roster, the variant matrix, captions, costs per variant, and the consent and disclosure work that keeps the whole thing defensible.

Why "real-looking" is the core requirement

UGC ads work because they look like a person filmed them on a phone. That sets a high bar for the avatar. A face that has been re-generated from scratch tends to come out smooth and evenly lit, which is exactly the "plastic" look viewers scroll past. We cover the causes in why AI avatars look plastic.

PitchTwin builds each avatar from a real video of the person and keeps that footage, so the result looks 1:1 like the creator: their skin texture, hair, room lighting and their own expressions. For UGC, that is the feature. More detail on the approach is on the avatar from video page.

Product design

The creator roster

Your generator needs people. Two common models:

  • Customer-supplied creators. The brand brings its own creators or staff, uploads their clips and manages consent.
  • A licensed roster. You contract a set of creators, pay them for the right to use their avatar in ads, and offer them inside your product.

Either way, each creator records one clip of 10 to 120 seconds on a phone. Encourage the setting your customers' ads will use: a kitchen, a car, a desk, natural light. A creator who wants two looks records two clips and gets two avatars. PitchTwin normalizes phone input (rotation, HEVC to H.264, long side up to 1920 px, 25 fps), so creators can upload straight from the camera roll. Ask for a quiet room, steady framing and even light on the face.

The variant matrix

Ad testing is combinatorial. Break each script into slots and let the user supply options for each:

SlotExample options
Hook"I did not expect this to work." / "Okay, real talk about my skin." / "Three weeks in, here is what changed."
BodyBenefit A / Benefit B
CTA"Link is below." / "Use code SPRING for 20% off."
CreatorAvatar 1 / Avatar 2 / Avatar 3

Three hooks, two bodies, two CTAs and three creators is 36 variants from a few minutes of writing. Your UI should make it easy to prune the matrix before rendering, because the cheapest variant is the one you never generate.

Script mode for pacing

UGC speech is punchy: short lines, small beats between them. Use script mode and put a short pause after the hook so it lands. Keep voice.speed near 1.0 or slightly above (the allowed range is 0.7 to 1.2).

Generating the matrix

Here is a minimal Python batch that renders every combination and records the generation IDs:

import itertools
import os
import requests

API = "https://pitchtwin.ai/api/v1"
HEADERS = {
    "Authorization": f"Bearer {os.environ['PITCHTWIN_API_KEY']}",
    "Content-Type": "application/json",
}

hooks = [
    "I did not expect this to work.",
    "Okay, real talk about my skin.",
]
bodies = [
    "The setup took me five minutes, honestly.",
    "It replaced three products in my routine.",
]
ctas = ["Link is below.", "Use code SPRING for 20% off."]
avatars = ["av_3f9c1e7a2b4d6e80", "av_8a21c0d4e5f61b37"]

jobs = []
for avatar_id, hook, body, cta in itertools.product(avatars, hooks, bodies, ctas):
    payload = {
        "avatar_id": avatar_id,
        "script": [
            {"text": hook, "pause_after_ms": 500},
            {"text": body, "pause_after_ms": 300},
            {"text": cta, "pause_after_ms": 0},
        ],
        "voice": {"speed": 1.05, "language": "en"},
        "webhook_url": "https://app.example.com/hooks/pitchtwin",
    }
    r = requests.post(f"{API}/generations", json=payload, headers=HEADERS)
    r.raise_for_status()
    jobs.append({"id": r.json()["id"], "avatar_id": avatar_id,
                 "hook": hook, "body": body, "cta": cta})

print(f"queued {len(jobs)} variants")

Store jobs in your database so each finished video can be tied back to its hook, body, CTA and creator. That mapping is what lets your users learn which hook wins.

Each POST returns 202 with status queued. When a render finishes, your webhook receives generation.succeeded or generation.failed, signed with the PitchTwin-Signature header. Verify it and handle failures.

Captions and finishing

Most feed video is watched muted, so captions are not optional. Every generation returns words with word-level timings, which is enough to render the bold, word-by-word captions UGC ads use. The approach is in captions from word timestamps.

Your finishing step, usually ffmpeg in a worker, can then:

  • Burn in captions from words.
  • Add a product shot or B-roll cutaway over the body section.
  • Reframe or crop for each placement your customers buy.
  • Add the end card and any required disclosure label.

The API delivers an MP4 without a watermark or vendor branding, so the finished ad belongs entirely to your customer. Links to generated files expire after 24 hours, so copy the MP4 into your storage before post-processing.

Bring your own voice performance

Some teams want more control over delivery: a specific emotional read, or a voice design from their own ElevenLabs account. Audio mode accepts your audio file (up to 10 minutes) with optional character alignment, and ElevenLabs /with-timestamps output is accepted as is. It is also the cheaper rate.

Costs per variant

PitchTwin charges prepaid tokens that never expire: 1 token per second of video with your own audio, 1.5 tokens per second when PitchTwin voices the script. For a 20-second ad variant:

PitchTwin voiceYour own audio
Tokens per 20 s variant3020
50 variants1,500 tokens1,000 tokens
Cost on Growth ($39 / 3,000)$19.50$13.00

Failed generations are refunded in full automatically. Since there are no seats or subscription, you can price your own product per variant or per credit without a fixed vendor bill underneath. See pricing for all packs.

A UGC generator multiplies one person's likeness across many ads, so consent needs more care than in most avatar products:

  • Specific scope. Creator agreements should name the uses (paid social ads for named brands or categories), the term, and whether the voice is cloned. Vague "any use" clauses are increasingly risky.
  • Per-brand approval. Let creators approve or block categories, so nobody ends up advertising something they would not endorse.
  • Revocation. When a licence ends, delete the avatar. DELETE /avatars/:id also deletes the cloned voice.
  • Disclosure. Ad platforms and some laws expect AI-generated or altered content to be labelled. Build a default label into your export.

PitchTwin's Terms require every person in an avatar to have agreed to it. For a broader look at the rules, read consent and likeness rights for AI avatars (not legal advice).

Try it

Record a short phone clip, create an avatar in the dashboard and generate a few hook variants in the Playground before writing any code. Sign up for 60 free seconds, no card required.

Keep reading