Seedance for ads: a practical guide

What Seedance does well for ad work

Seedance is ByteDance's video model that generates picture and sound together from text, image, video and audio inputs; ByteDance Seed calls it a "unified multimodal audio-video joint generation architecture" (ByteDance Seed). For ad work the useful parts are long single clips, many reference inputs, and native audio.

Seedance is sold three ways: as an API on BytePlus ModelArk, as Volcengine Ark in China, and inside ByteDance's Dreamina app. The current flagship is Dreamina Seedance 2.5, and the API docs carry the banner "Dreamina Seedance 2.5 is now fully available" (BytePlus docs).

Two of these matter most for ads. A 30-second single generation on 2.5 can hold a full short ad beat sequence in one clip. Audio-only reference input on 2.5 means you can supply a finished voice track and generate picture to it.

Which ad formats Seedance suits

Seedance suits formats where one person or one product has to stay the same across shots and where the speech has to match the picture. The table below is guidance from the published capabilities, not test results.

For animated formats such as Pixar-style ads, the same reference-image mode keeps a stylised character on model across shots. Feed it a character sheet rather than a single pose.

Workflow: from script to finished Seedance ad

Start from a locked script, cut it into shots, build reference stills, generate each shot with audio, then finish in an editor. Seedance makes clips; the ad is assembled after.

1. Split the script into beats. Hook, problem, product, proof, call to action. Give each beat a target duration and one spoken line at most.
2. Write a shot list. One row per shot: duration, framing (selfie, two-shot, packshot), the line of dialogue, and the aspect ratio. Keep each shot inside the clip limit of the model you use: 15 seconds on 2.0, 30 on 2.5.
3. Build reference stills. Make a presenter image (front-on, neutral light), a clean product packshot and one location still. Reuse the same files for every shot so the person and product match.
4. Decide where the voice comes from. Either write the line into the prompt and let Seedance generate speech, or record or generate the full voice track first and pass it as reference audio. On 2.5 the second route keeps the read consistent across cuts.
5. Generate each shot. Use image-to-video with the presenter and product references. Generate a few takes per shot and pick on mouth sync, hands and product label legibility.
6. Edit. Join the shots, add burned-in captions, music and a final card. Put prices, offers and claims in as text in the editor, not in the prompt, so they are exact and easy to change. The generated audio is mono, so mix voice and music in the editor.
7. Export per placement. Seedance has no 4:5 option, so for a 4:5 feed placement generate 3:4 or 9:16 and crop in the editor.

A 30 to 45 second ad on Seedance 2.0 is typically three or four clips. On 2.5 you can try the whole hook-to-product sequence in one 30-second generation, then extend or cut in extra shots.

For a variant, change one element per version: the hook line, the presenter, or the setting. Hold the body and the call to action fixed so you know what changed.

Seedance prompt structure for ads

A Seedance ad prompt works best in a fixed order: subject and reference, setting, action, camera, spoken line, sound. Fill the placeholders and delete anything the reference images already show.

The templates below are written for this guide. Square-bracket items are placeholders.

UGC selfie hook (9:16, image-to-video with presenter reference)

Vertical selfie video, handheld front camera at arm's length.
[presenter from reference image] in [location, e.g. small kitchen, morning light].
She looks straight into the lens, leans in slightly and says:
"[hook line, one sentence]"
Natural pauses, slight camera shake, no music, room tone only.
Duration [5-8] seconds.

Product hero shot (first and last frame)

First frame: [product packshot reference] on [surface] against [background colour].
Last frame: same product, label facing camera, centred, filling the middle third.
Slow push-in from medium shot to close-up, soft key light from the left.
[one physical action, e.g. condensation runs down the can / lid lifts].
Sound: [single sound effect], no voice.
Leave clear space at the top for on-screen text.

Podcast two-shot (reference audio drives the picture)

Two people at studio microphones with headphones, [studio or living-room set].
Left: [host A reference]. Right: [host B reference].
Medium two-shot, static camera, warm practical lights in background.
Lip sync to the reference audio: host A asks, host B answers.
Small natural gestures, eye contact with each other, not with camera.

Animated explainer scene

3D animated style, [character from reference sheet] in [setting].
The character points to [object representing the problem], then picks up [product].
Camera: slow orbit left, ends on the character holding [product] to camera.
Character says: "[one-line explanation]"
Bright, soft lighting, [colour palette].

Brand-film establishing shot (21:9 or 16:9)

Wide establishing shot of [location] at [time of day], [weather].
[Small human element, e.g. a cyclist crossing the frame].
Slow aerial drift forward, no cuts.
Ambient sound: [wind / traffic / sea], no dialogue.
Colour grade: [look], film grain.

Limits, terms and policy notes

The main limits are licensing and access, not picture quality. Which Seedance route you use decides what you are allowed to do with the output.

Dreamina app. The Dreamina terms say "Our Services are generally provided for private, non-commercial use", with commercial use only "if permitted" under additional terms (CapCut, Dreamina Terms of Service). Read the plan terms that apply to you before using Dreamina output in a paid ad.

BytePlus API. No explicit "you own the output" sentence was found in the Model Services terms. The terms do include an IP indemnity: "BytePlus shall defend you against a third party claim alleging that your permitted use of any Output generated by a BytePlus Model Services ... infringes the intellectual property rights of that third party" (BytePlus terms). The same terms state: "The BytePlus Model Services are not available in the United States." Section 4.2.3 also restricts resale into UGC or AI-creation platforms.

Pricing, as published. The BytePlus API bills per token: "Estimated video price: Token unit price × Token consumption" (BytePlus pricing). The rates listed, in USD per million tokens to match the page's worked examples, were:

The price table on that page did not parse cleanly when checked, so confirm these rates on the live page before you budget with them. The page also says: "You are only charged for successfully generated videos." To turn on Seedance 2.0 or 2.5 you need a balance above USD 30, a USD 30+ AI Savings Plan, or a Seedance resource pack (BytePlus docs).

BytePlus also sells Seedance 2.5 plans named Light, Production and Premium, "valid for 3 months"; their prices did not render when checked (BytePlus product page).

Dreamina's pricing page showed prices in Philippine pesos, localised to the location it was fetched from, so a USD price is not published to us. As shown: Basic PHP 910 a month (9728 credits), Standard PHP 2190 a month (23849 credits), Advanced PHP 5,803.00 for the first month then PHP 9,672.00 a month (108891 credits), Ultra PHP 18,979.00 for the first month then PHP 31,632.00 a month (365031 credits), plus free daily credits (Dreamina pricing). Your page will likely show your local currency.

Platform labels. Meta labels ad images "created or edited using third-party AI tools with AI info" (Meta). TikTok allows AI-generated ad content only if you apply the AIGC label or your own clear disclaimer, and states "your ad will be rejected or restricted" if undisclosed AI content is found (TikTok ads policy). Details are in AI-generated content labels on Meta and TikTok ads.

Faces and claims. TikTok's ad policy does not allow misuse of a public figure's likeness. Build presenters from generated references, not photos of real people, and keep health, money or results claims out of the dialogue unless the brand has approved them.

Seedance in a one-script, every-format workflow

Seedance fits the one-script method because the same reference stills and voice track can feed several formats. One script can become a selfie version, a podcast two-shot and a 21:9 brand cut without rewriting the core lines.

The order that keeps this manageable: lock the script, record or generate the voice once, then change only the picture per format. The full method is in one winning script, made in every ad format.

Anglr pays AI creators, editors and agencies on performance to remake winning ad scripts in every format. Anglr provides the scripts and angles; the creator makes the versions, and every video gets a QC and compliance review before it goes live. If you already make full ads with tools like Seedance, see how it works for creators.