UGC ad video

Produce a short, creator-style ad: a presenter talking straight to camera (testimonial, demo, unboxing) or a faceless voiceover over product footage. The output reads like organic social content, not a polished broadcast spot. The levers are a talking-avatar presenter, the script or audio that drives it, and the background/framing/captions tuned for the target feed.

Inputs to collect

• The format: direct-to-camera testimonial, product demo, unboxing, or faceless voiceover. This decides whether you need a presenter on camera at all.
• The script (for the TTS path) or your own audio file (for the lip-sync path). One or the other, never both.
• The product and the single claim the ad should land. (Ask for real specifics. Do not invent features.)
• Target platform and orientation: 9:16 for TikTok / Reels / Shorts, 16:9 for YouTube / web.
• Optional: a chosen avatar/voice, a brand background, whether captions should be burned in.
• Consent for a real person. If the ad uses a real presenter, voice, or likeness, confirm the user owns it or has permission. Do not impersonate a real or public figure without authorization, and disclose synthetic media where it could be mistaken for real.

Models

• Presenter on camera. Default: HeyGen Avatar V (heygen:avatar@5). A registered-look talking avatar with two drive modes (script-to-speech or your own audio) plus background, fit, and caption controls. Best general pick for a person reading the ad.
• Presenter alternative: OmniHuman-1.5 (bytedance:5@2) when you want to animate a specific supplied photo into a speaking presenter rather than pick from a registered catalog.
• Faceless voiceover: generate the spoken track with the voiceover skill (a TTS model), then lay it over product footage produced by product-ad-video. No avatar needed.
• Confirm each model is live and inspect its schema via runware-models + runware-run before calling. Never hardcode a stale choice, since the catalog changes weekly.

Workflow

This is an async outcome (video). Follow the runware-run contract.

1. Resolve the presenter model's schema (runware-run) and confirm the field names: the required avatar/look input, the speech block, the inputs.audio field, and the settings for background/fit/caption.
2. Upload any inputs (a supplied photo for OmniHuman, a brand background image, or your own audio) and reference them by URL or UUID.
3. Build the request with taskType: "videoInference", the model AIR, the avatar/look, and exactly one drive path (see Technique).
4. Set width/height for the target orientation and the settings for background, fit, and captions.
5. Run async: the task returns a taskUUID. Poll getResponse until terminal, then read videoURL from the result.
6. Iterate on a short take first (a 10-second script renders fast and surfaces avatar/voice mismatches), then send the full cut.

Technique

Ground the ad in a named build-order so it earns attention before it asks for anything:

1. Hook. The first two seconds: a problem, a bold claim, or a pattern interrupt. Attention drops off fast on social, so the hook carries the ad.
2. Context. Who this is for and the problem it solves.
3. Product moment. Show or name the product doing the thing.
4. Proof. A real demonstration or genuine result. Never a fabricated testimonial, metric, or review.
5. Call to action. One clear next step.

Fill this in with the user's real copy before generating the script, then flatten it into speech.text (or the recorded audio):

HOOK (0-2s): <problem, bold claim, or pattern interrupt>
CONTEXT: <who this is for and the problem it solves>
PRODUCT MOMENT: <the product doing the one thing>
PROOF: <a real demonstration or genuine result, supplied wording only>
CALL TO ACTION: <one clear next step>

Load references/examples.md for three worked recipes (TTS testimonial, own-audio demo, 9:16 captioned cut) with the full request and result shape.

Pick exactly one drive path. Sending both errors.

• TTS path: supply a speech block with text (the literal script) and voice (the catalog voice). Best for fast copy iteration and localization. Voice tuning (speed, pitch, volume, language) only exists on this path.
• Lip-sync path: supply inputs.audio (a public URL or an uploaded-asset UUID) and omit the speech block entirely. Best when you already have the exact delivery: a real human recording, a voice clone, or a track from a separate TTS run (the voiceover skill). The model extracts phonemes from the audio and animates the mouth to match.

Compose for the feed, not the model. Most of what makes UGC read as native happens around the presenter:

• Background: run settings.removeBackground: true together with either settings.backgroundColor (a hex string) or inputs.background (an image). Removing the background alone just falls back to the avatar's original environment. Matting only works for avatars trained with it enabled.
• Orientation first, then resolution. Choose 9:16 vs 16:9 by where the ad plays, then pick the resolution. Match any background image's aspect ratio to your output dimensions.
• Fit: for portrait output, settings.fit: cover is almost always right (it crops the landscape source to fill the tall frame). contain letterboxes and rarely looks production-ready in portrait.
• Captions: settings.caption: true burns subtitles into the video. Use it for silent-autoplay feeds and accessibility. The style is fixed, so render without it and overlay your own if you need custom typography.

Persona match. Pair the avatar and voice deliberately. A mismatched voice and face is jarring and breaks the organic feel UGC depends on. Lock both before iterating on copy.

Parameters that matter

• Drive path is mutually exclusive: speech (TTS) or inputs.audio (lip-sync), never both.
• speech.speed: 0.5 to 1.5, default 1.0. Faster reads suit short social cuts where attention drops early.
• speech.pitch: -50 to +50, default 0. Keep adjustments small (±5 to ±15) or it sounds processed.
• speech.volume: 0.0 to 1.0, default 1.0. Lower it to leave room for a music mix.
• speech.language: a BCP 47 locale (en-US, es-ES, and so on). Same voice, translated script, one locale code per market is the cheapest way to localize at scale.
• width/height: fixed 16:9 or 9:16 sets (e.g. 1080×1920 portrait, 1920×1080 landscape). 1080p is the production default, 720p for iteration.
• settings.removeBackground, settings.backgroundColor, inputs.background, settings.fit, settings.caption: the composition levers above.
• Confirm exact field names against the live schema (runware-run); never guess.

Quality bar

• The hook lands in the first two seconds and the build-order flows Hook → context → product moment → proof → call to action.
• Exactly one drive path is set. The request validates (dry-run via runware-run if unsure).
• No fabricated testimonials, metrics, reviews, or on-screen claims. Only real, supplied wording appears as readable copy.
• Avatar and voice are a believable pair, not a jarring mismatch.
• Orientation, fit, background, and captions match the target platform: portrait plus cover plus burned-in captions for Reels/TikTok, landscape for YouTube/web.
• The cut reads as native creator content, not a stiff corporate spot.

Related skills

runware-run, runware-models, runware-prompting; voiceover (generate the spoken track for the lip-sync path), talking-avatar (presenter mechanics in depth), product-ad-video (footage for faceless cuts).