Seedance 2.0 Video Prompts: How I Make AI UGC Ads That Look Real

Why UGC Format + AI Video Is the Unlock

UGC-style ads — a normal-looking person talking to camera about a product — have outperformed polished brand ads on Meta for years. The bottleneck was always production: finding creators, briefing them, waiting a week, paying $150-400 per video, and getting one usable take out of three.

Seedance 2.0 removes the bottleneck. I can generate a 10-15 second clip of a believable person — a mid-40s professional at his desk, a guy on a tailgate, an exec sneaking a lunch break — for a couple dollars of compute, in minutes. That means I can test eight different avatars against eight different angles in an afternoon, which no human production pipeline on earth can match. In 2026, creative volume wins on Meta, and this is how you get volume.

The Prompt Structure That Works (60-100 Words)

Lazy prompts produce lazy video. After hundreds of renders, this is the five-part structure I use for every single clip:

1. Subject, specifically. Not "a man." Instead: "a 45-year-old man with greying hair, casual button-down, sitting at a home office desk, speaking directly to a phone camera propped in front of him."
2. Timestamped shots, one camera move each. "0-4s: static medium shot as he leans in. 4-8s: slow push-in to close-up. 8-10s: he holds up one finger, camera static." One move per shot. This is the single biggest realism lever.
3. Lighting, explicitly. "Soft window light from the left, warm practical lamp in background bokeh." No lighting direction = flat plastic default light = instant AI look.
4. A style anchor. "Shot like a casual iPhone selfie video" for UGC, or "Christopher Nolan cinematic" / "Apple keynote aesthetic" for b-roll. The anchor sets a thousand small decisions at once.
5. A hard negative. "ABSOLUTELY NO text, letters, words, UI, screens showing content, holograms, or floating panels."

What Breaks — Every Time

Save yourself the failed renders. These are the consistent failure modes:

• On-screen text. Any request for readable words, phone UI, or laptop screens produces garbled alien script. Always. Add text in post, never in the prompt.
• Two camera moves in one shot. "Pan while zooming as she turns" gives you seasick, uncanny motion. One move per timestamped beat.
• Overloaded action. Asking a character to walk, gesture, pick something up, and react in 10 seconds produces rubber-limbed chaos. One action per beat, same as camera.
• Generated audio on non-speech clips. I generate video silent and add voiceover in post — cloned or recorded VO synced in the edit beats whatever ambient audio the model invents.
• Vague vibes. "A cool video of a guy loving a product" is not a prompt. Specificity is the entire skill, same as vibecoding — the model can't guess what you didn't say.

The Realism Checklist

What separates "obviously AI" from "scrolled past it, thought it was real":

• Ordinary people, not beautiful people. Prompt for slightly imperfect, age-appropriate faces: "looks like a real person, not a model, subtle skin texture." Gorgeous symmetrical faces scream AI.
• Ordinary settings. Messy desk, car interior, kitchen counter, garage. Pristine minimalist sets read as fake because real UGC isn't shot on sets.
• Phone-camera framing. "Vertical 9:16, framed like a propped-up phone, slightly imperfect framing" — perfect tripod symmetry is a tell.
• Natural light behavior. Window light, lamp bokeh, slight overexposure. Studio-even lighting is the plastic look.

The goal isn't a beautiful video. The goal is a video that looks like a customer made it. Ugly-on-purpose beats cinematic for UGC every time.

From Clip to Ad: The Post Pipeline

A raw Seedance clip is footage, not an ad. My pipeline: generate the talking-head clip silent → generate or record the voiceover separately → sync in the edit → add captions (word-by-word captions lift retention meaningfully) → cut in 2-3 seconds of b-roll if the script references anything visual → end card with the offer. Total edit time per ad: 20-30 minutes once the pieces exist.

The script matters more than the render. The formula that converts: a specific person names a specific pain ("I spent $2,000 on AI courses that were recycled YouTube videos"), a turn ("then I found..."), 2-3 concrete specifics about the product, and a low-friction close. The actor is a user telling their story, never the author pitching. If you're building a content engine around this, my guide to automating content creation with AI covers the surrounding system, and the 2026 AI video generator test shows how Seedance stacks against the field.

Testing at Volume

Because each video costs almost nothing, the strategy changes: don't perfect one video, generate one concept across multiple avatars. Same script beats, different person: the 50-something IT veteran, the guy in the grey hoodie, the blue-collar tailgate guy. Then let Meta's delivery tell you which avatar your market trusts — in my account, two specific "regular guy" avatars produced purchase signal while the polished-looking ones did nothing. You cannot learn that insight without volume, and you can't afford volume with human actors.

One rule I never skip: watch every render fully before it goes anywhere near an ad account. Check hands, check teeth, check background objects, check that nothing morphs mid-shot. One in five renders has a flaw you'll only catch by looking.

---

FAQ

What is Seedance 2.0 and why use it for UGC ads?

Seedance 2.0 is ByteDance's video generation model, and in 2026 it's one of the strongest for realistic people talking to camera. For UGC-style ads it means testimonial-format creative — a person at a desk, in a car, on a lunch break — without hiring actors, booking shoots, or paying creators per video.

Why do AI-generated videos look fake, and how do you fix it?

Three main causes: over-directed prompts (multiple camera moves per shot), missing lighting direction (the model defaults to flat plastic light), and asking for on-screen text or UI, which always garbles. Fix with one camera move per timestamped shot, explicit lighting, a style anchor, and a hard negative prompt banning text and screens.

Can I put text or app screens in AI video prompts?

No. Every 2026 video model garbles rendered text, phone screens, and UI. Keep prompts purely cinematic — people, environments, light, motion — and add headlines, captions, and screen recordings in post-production.