Blog
/
AI Video Ads
AI Video Ads

Seedance 2.5 for UGC ads: prompts, 30-second one-take ads and settings that work

The best Seedance 2.5 prompts for UGC ads read like a shot plan: the camera, its flaws, what must stay fixed, and timecoded beats with dialogue. Two structures work well, a 30-second 10-cut selfie vlog or a single unbroken take. In Arcads, an AI video ad platform, Seedance 2.5 makes 4–30 second clips with audio.

Updated November 19, 2026 · By the Arcads team. Arcads is an AI video ad platform for making UGC-style and AI actor video ads.

Key facts at a glance

Seedance 2.5 is ByteDance's video model and the default video model in Arcads. These are the facts that matter for UGC ads:

  • When did Seedance 2.5 arrive on Arcads? August 7, 2026, with 72 hours of unlimited use for Arcads users (Source: @arcads_ai and @rom1trs on X, Aug 7, 2026)
  • Clip length: 4 to 30 seconds, with audio. It's the default video model in Arcads (Source: Arcads Help Center, "AI Video Models in Arcads")
  • Standard cost: 210 credits for 5 s at 720p, 520 credits for 5 s at 1080p (Source: Arcads Help Center, "How does Arcads pricing work?")
  • Cheaper previews: Draft mode renders at 480p (18.75 credits/s). "Generate final version" re-renders the same take at 1080p (104 credits/s) (Source: Arcads Help Center, "Seedance 2.5 Draft mode")
  • Where it runs: Studio, Workflows, Mark (the Arcads agent) and the Arcads MCP in Claude and ChatGPT (Source: Arcads Help Center)
  • AI actors built on it: Actor+, launched August 14, 2026, is "powered by Seedance 2.5" (Source: @arcads_ai on X, Aug 14, 2026)

Is Seedance 2.5 good for UGC ads?

Yes. Seedance 2.5 suits UGC ads because it generates audio with the picture, reaches 30 seconds in one take and follows long, structured prompts. Arcads lists it as the default for "most ads, with audio, up to 30 seconds". UGC suits it for three reasons. The audio comes out in the same pass as the picture. The 30-second ceiling fits a whole hook-to-CTA ad in one generation. And it follows long, structured prompts closely enough that you can direct camera faults on purpose. Those faults are what make a clip read as self-filmed instead of produced.

Arcads tested it in public before launch. On July 31, 2026, Arcads co-founder Romain Torres posted "Seedance 2.5 vs. Seedance 2.0 for AI UGC" with the note "Coming soon on Arcads". A POV-video experiment followed, then a sitcom test on August 3. The model went live on August 7. A week later, Arcads launched Actor+, its AI actor library built on Seedance 2.5.

Where can you find real Seedance 2.5 prompt examples?

Arcads and its co-founder Romain Torres (@rom1trs) posted complete Seedance 2.5 prompts on X between August 12 and August 24, 2026, each next to the video it produced. The public examples cover two structures:

  • Aug 12: Public example on X: Late-night gym vlog; Structure: 30 s, 10 cuts, self-filmed
  • Aug 12: Public example on X: Late-night cooking vlog; Structure: Self-filmed, landscape
  • Aug 13: Public example on X: Dawn-run vlog; Structure: Self-filmed, vertical, Mini-DV look
  • Aug 14: Public example on X: Flash-mob music video; Structure: Single unbroken 30 s take
  • Aug 17: Public example on X: House party to pool dive; Structure: Single unbroken 30 s take
  • Aug 17: Public example on X: Emotional close-up; Structure: Single unbroken 15 s take
  • Aug 24: Public example on X: Dance from a drawn storyboard; Structure: 12-shot timecoded storyboard + character sheet

All of these prompts share one idea: they say exactly how the footage is imperfect.

How do you write a 10-cut selfie vlog prompt for Seedance 2.5?

A 10-cut Seedance 2.5 vlog prompt uses eight labeled sections: camera, camera-logic rules, look, style, character and setting, continuity rules, a 30-second storyboard, and a final instruction. The public gym, cooking and dawn-run prompts use them in this order:

  1. CAMERA: the device and its flaws. The gym prompt opens "DV 16mm tape camcorder handheld aesthetic, landscape framing". It asks for "realistic hand shake, slightly crooked framing, delayed autofocus, awkward zoom-ins and zoom-outs". The running version adds shake "amplified by her stride". The cooking version adds a "brief lens steam-up over the hot pan that clears naturally".
  2. Camera-logic rules: the line that stops the most common AI-UGC giveaway. Every vlog prompt includes "never show her placing, adjusting, or setting up a camera". It also asks that cuts to a fixed angle behave "as if the camera had already been positioned before the shot began".
  3. LOOK: the tape character, highlight bloom, exposure flicker and real skin texture.
  4. STYLE: the tone in plain words, and what it must not become ("not a polished commercial").
  5. CHARACTER and SETTING: one person and one place, each described once, in detail.
  6. CONTINUITY RULES: a list of things that must not change (face, hair, outfit, props) and things that must not appear (extra people, duplicated limbs, a camera in the mirror).
  7. STORYBOARD: "30 seconds total, 10 cuts", with roughly 3 seconds per beat. Each beat names its framing ("arm's-length selfie", "fixed external medium shot") and gives a short line of dialogue.
  8. FINAL INSTRUCTION: a one-paragraph restatement of the goal.

Template: 10-cut UGC product vlog (adapt and test)

The template below uses the same section order for a skincare product ad. It's a starting point, not a tested prompt, so generate it in draft mode first.

CAMERA: Handheld phone-camera aesthetic, vertical 9:16. A woman in her early thirties films herself in selfie mode in her bathroom in the morning. Keep natural hand shake, slightly off-center framing, autofocus briefly hunting between her face and the product, one awkward push-in. Mix handheld selfie shots with two fixed shots from the counter. Never show her placing or setting up the phone. When cutting to a fixed shot, jump-cut as if the phone was already there. The phone never appears on screen or in the mirror.
LOOK: Soft window light, realistic skin texture with visible pores, no beauty filter, mild exposure flicker when she moves toward the window.
STYLE: A real "get ready with me" clip, relaxed and a little rushed. Not a commercial.
CHARACTER: [age, hair, clothing in one sentence]. Same person, same outfit in every shot.
PRODUCT: Use @product exactly as in the reference: same bottle shape, label and cap color in every shot. The label faces the camera when she holds it up.
CONTINUITY RULES: No face, hair or outfit changes. No extra people, including in the mirror. No text on screen. No duplicated hands.
STORYBOARD: 30 seconds total, 10 cuts, about 3 seconds each.
1 (selfie) She yawns at the camera. Dialogue: "[hook line, under 8 words]"
2 (selfie, close) She shows her skin problem. Dialogue: "[problem line]"
3 (fixed counter shot) She picks up @product. No dialogue.
4 (selfie, close on hands) She applies it. Dialogue: "[how she uses it]"
5-8 [demo, texture close-up, reaction, one honest caveat]
9 (selfie) She checks the mirror and smiles. Dialogue: "[result line, no medical claims]"
10 (selfie) She grabs her keys. Dialogue: "[CTA line]"
FINAL INSTRUCTION: It must feel self-recorded, with believable pacing and small imperfections. Keep the product identical throughout.

Save your product as an Element first (Studio → Asset library → Elements, up to 8 reference images) so @product keeps the label and packaging stable from shot to shot.

How do you write a 30-second one-take prompt for Seedance 2.5?

A one-take Seedance 2.5 prompt says "single unbroken take, 30 seconds" up front, gives one timecoded line per beat describing what the camera does, and ends with a lock on wardrobe and a list of banned artifacts. The public pool-dive and flash-mob prompts work this way and drop cuts altogether. The pool-dive prompt opens: "Single unbroken handheld take throughout, 30 seconds total". It then gives one timecoded line per beat (0–4s, 4–8s, 8–13s and so on), and every line says what the camera does at that moment. One example: "Camera smoothly rotates from following behind her to facing her". It ends with a lock on wardrobe and a list of banned artifacts: "No text overlays, no flickering, no ghosting, no morphing artifacts... no hard cuts at any point."

The 15-second emotional close-up shows a second trick: freeze everything that shouldn't change. Its premise says the signs of crying are "ALREADY PRESENT IN THE FIRST FRAME and stays constant". It then lists the only things allowed to move: "facial muscle position, tear fluid, breathing movement, hair position, cloth position, and wet gloss on skin". If a product label or a skin tone drifts during your one-take, state what is fixed and what can move in the same way.

The flash-mob prompt adds a third idea. It treats the soundtrack as a finished studio track "laid directly over the footage in the edit". The music doesn't play inside the scene. Use that when the music matters more than room sound.

Template: one-take product demo (adapt and test)

One unbroken handheld take, 30 seconds, vertical 9:16, no cuts, no scene changes.
FIXED FOR THE WHOLE CLIP: the same man, late twenties, grey hoodie; @product (matte black water bottle) with its logo always readable; overcast daylight; a city park path.
ALLOWED TO CHANGE: his walking pace, hand position, facial expression, wind in his hair.
0-4s: Close selfie while walking. He says: "[hook]". Slight bounce from his steps.
4-10s: He lifts @product into frame and turns it so the logo faces the lens. The camera stays at arm's length.
10-18s: He opens the lid with one hand and takes a sip without stopping. Audio: the lid click, park ambience.
18-25s: He tilts the phone down to show the bottle clipped to his bag, then back up to his face. He says: "[benefit line]".
25-30s: He stops walking, holds @product next to his face and says: "[CTA]". He holds the frame still for the last second so captions can sit on it.
No text overlays, no logo distortion, no morphing, no hard cuts, natural skin, realistic hand shake throughout.

Which Seedance 2.5 settings should you use in Arcads?

For UGC ads in Arcads, use Seedance 2.5 with draft mode on for the first pass, 9:16 at 4–30 seconds, Elements for the face and product, then "Generate final version" at 1080p for the keeper.

  • Model: What to pick: Seedance 2.5 (default); Why: Built for "most ads, with audio, up to 30 seconds", per Arcads
  • Duration: What to pick: 4–30 s. Use 30 s for one-take ads and 10–15 s for single beats you will stitch; Why: Seedance 2.5 range on Arcads
  • Draft mode: What to pick: On, for every first pass; Why: 480p preview at 18.75 credits/s. The resolution is fixed at 480p while it's on
  • Final version: What to pick: "Generate final version" on the keeper; Why: Same prompt, references, length and aspect ratio, rendered at 1080p as a new video
  • Resolution (no draft): What to pick: 720p to test hooks, 1080p for delivery; Why: 210 vs 520 credits per 5 s at standard rates
  • References: What to pick: Elements (@character, @product) or a start frame; Why: Keeps faces and packaging stable across takes
  • Aspect ratio: What to pick: 9:16 for Reels/TikTok, plus a 4:5 version for Meta feed; Why: Resize Video in Arcads is free (0 credits)

How much does a 30-second Seedance 2.5 ad cost?

At Arcads' standard rates, a 30-second clip costs:

  • Draft (480p): 30 × 18.75 = about 563 credits
  • 720p: 30 × 42 = 1,260 credits (from 210 credits per 5 s)
  • 1080p: 30 × 104 = 3,120 credits (from 520 credits per 5 s)

So a draft that you throw away costs about 18% of a 1080p take that you throw away. Romain Torres described the same habit in a public September 2, 2026 post about an AI Gucci spec ad: "480p first, keepers at 1080". Arcads notes that rates can vary by account. The Generate button always shows the real cost before you run it.

Can Seedance 2.5 do image-to-video?

Yes. Seedance 2.5 can animate a still image in Arcads: in Studio, generate or upload an image, then click Video on it to animate it with a video model. Arcads' own model guide chains it this way: image first, then video, then a talking actor layered on top, then captions. For motion-design ads made from a still, Arcads published a Claude skill on August 10, 2026 ("drop in any still image, it writes the prompt as timed beats"). That skill is now image-to-motion in the Arcads skills library, and it uses Seedance 2.5. A public dance example posted on X on August 24 goes further. Claude writes a timecoded 12-shot storyboard, GPT Image 2.0 draws each shot plus a character sheet, and Seedance 2.5 renders each shot.

When did Seedance 2.5 come to Arcads?

Seedance 2.5 went live on Arcads on August 7, 2026, after public tests from July 31. Draft mode and the Reality TV ad skill followed later in 2026.

  • Jul 31 – Aug 3: Public tests: 2.5 vs 2.0 for AI UGC, POV clips, a sitcom test ("Coming soon on Arcads")
  • Aug 7: Seedance 2.5 live on Arcads, with 72 hours of unlimited use
  • Aug 10: Claude skill for Seedance 2.5 motion design (now image-to-motion)
  • Aug 12–24: Full prompt posts: vlogs, one-takes, music video, drawn storyboard
  • Aug 14: Actor+ AI actor library, "powered by Seedance 2.5"
  • By November 2026: Draft mode: 480p preview, then "Generate final version" at 1080p. Also available in Mark and the MCP
  • Nov 16: Reality TV ad skill: product reference + Seedance 2.5 or Wan, 30-second 1080p output (see the guide to AI reality TV ads)

Can you use Seedance 2.5 in Claude, ChatGPT or Mark?

Yes. Seedance 2.5 runs in Claude and ChatGPT through the Arcads MCP, and in Mark, the agent inside Arcads, all on the same Arcads account and credits.

  • Claude or ChatGPT: connect the Arcads MCP (https://mcp.arcads.ai as a custom connector in Claude, or the Arcads plugin in ChatGPT). The MCP has a Draft option on Seedance 2.5 generation and a separate tool to generate the final version. Ask for the draft explicitly, review it, then ask for the final.
  • Mark: tell Mark to render in Seedance 2.5 draft mode first. Arcads' help center says Mark will do this when you ask for it specifically.
  • Studio: toggle Draft mode in the Seedance 2.5 options.

A request you can paste into Claude or ChatGPT once Arcads is connected:

Using Arcads, turn my brief into a Seedance 2.5 prompt with labeled sections: CAMERA, FIXED, ALLOWED TO CHANGE, timecoded beats, banned artifacts. Show me the prompt first. Then generate one 30-second 9:16 draft in draft mode. Don't generate the 1080p final until I approve the draft.

Why do Seedance 2.5 UGC clips look fake, and how do you fix it?

Most Seedance 2.5 UGC failures come from six causes: the camera appearing in shot, labels drifting, unwanted cuts, a polished look, dialogue that runs long, and claims that break ad policy.

  • The camera appears in the shot or mirror. Add the camera-logic rules from Structure 1, and say "including in mirror reflections".
  • The product label changes between shots. Use an Element with several angles of the packaging, and list the label under FIXED.
  • The one-take cuts anyway. Repeat "no cuts, no scene changes" at the top and "no hard cuts at any point" at the end, as the pool-dive prompt does.
  • The clip looks too polished. Name the device (phone, DV camcorder) and give two or three specific flaws, not the word "realistic".
  • The dialogue runs long. Keep each line short enough for its beat. In the 10-cut vlogs, every line fits about 3 seconds.
  • Policy check. Arcads' prohibited-content rules ban sexually suggestive content and misleading claims. Keep characters and lines brand-safe, and don't let a "testimonial" make claims your product can't support.

Frequently asked questions

How long can a Seedance 2.5 video be on Arcads?

Between 4 and 30 seconds per generation, with audio. For longer ads, generate several clips and join them with Stitch Videos, which costs 0 credits.

How much does Seedance 2.5 cost on Arcads?

At standard rates, 210 credits for 5 seconds at 720p and 520 credits for 5 seconds at 1080p. Draft mode previews cost 18.75 credits per second at 480p. Your account's real cost is shown on the Generate button.

What is Seedance 2.5 draft mode?

A 480p preview of a Seedance 2.5 video. If you like the take, choose "Generate final version" and Arcads renders the same prompt, references, length and aspect ratio at 1080p as a new video. It's available on paid plans in Studio, Mark and the Arcads MCP.

How do I write a Seedance 2.5 prompt for a 30-second one-take ad?

Open with "one unbroken take, 30 seconds, no cuts". List what stays fixed and what can change. Give one timecoded line per beat that says what the person and the camera do. End with a list of banned artifacts such as morphing, flicker and hard cuts.

Can I use Seedance 2.5 in Claude or ChatGPT?

Yes. Connect the Arcads MCP in Claude or install the Arcads plugin in ChatGPT, then ask for a Seedance 2.5 generation. The MCP supports draft mode and has a tool to generate the final 1080p version.

Is Seedance 2.5 the best model for every ad?

No. Arcads recommends it as the default for most ads with audio. For other jobs, such as fast iteration, keyframed motion or editing existing footage, see the comparison of the best AI video models for ads.

Key takeaways

  • The strongest public Seedance 2.5 prompts describe the camera's flaws and the camera's logic, not just the scene.
  • Use the 10-cut vlog structure for testimonial-style UGC and the single-take structure for one continuous moment.
  • Say what is fixed and what is allowed to change. It's the simplest fix for drifting faces and labels.
  • Draft at 480p (about 563 credits for 30 s), then render only the keepers at 1080p (3,120 credits for 30 s at standard rates).
  • Save products and characters as Elements so every take uses the same references.

Keep reading: Claude skills that make ads · How to write a video ad script · How to create UGC ads

Sources

Arcads MCP

Top AI models, one prompt away in Claude, ChatGPT and Grok Bot.

Connect the Arcads MCP once, then generate video and images with top models from your AI assistant.

Add Arcads MCP