How to use Seedance for ads: a guide built from creator tutorials

What is Seedance, and how do creators use it for ads?

Seedance (Dreamina Seedance) is a video model sold to developers through BytePlus ModelArk. Most creators reach it through apps such as Higgsfield, MakeUGC, Arcads or CapCut. It takes text plus reference images, video and audio, and it generates speech and sound with the picture.

For ad makers, the useful part is reference control. You upload a character image, a product image and a location, name each one in the prompt, and the model holds them through the shot. Ian Ramírez Rivera explains why he moved a client ad to Seedance 2.5: "It treats them as constraints, not suggestions."

Facts from BytePlus's own prompt guide, checked 12 Oct 2026 (BytePlus ModelArk):

• Images are referenced by upload order as Image 1, Image 2 and so on.
• A prompt can take at most three reference video clips, with a combined length of no more than 15 seconds.
• Output becomes less stable when more than four people are referenced.
• The guide recommends one camera movement per shot.

Which tools do creators pair with Seedance?

Many tutorials pair an LLM that writes the prompts with an image model that makes the references, then use Seedance for motion and an editor for the finish. Based on 23 public tutorials posted between 12 Feb 2026 and 11 Oct 2026, reviewed by Anglr.

Claude shows up in about half of the tutorials. The pattern is the one in our Higgsfield MCP guide: the LLM plans and writes, and the video model only renders.

How do you make an ad with Seedance, step by step?

The tutorials agree on the order: references, then prompts, then batches, then the edit. They disagree on timing marks and on what kind of character sheet to use, and in both cases BytePlus's guide sides with the minority.

1. Make the stills before any video. Eugenio Zabell, Abhinav Kalyan, Aswad Tahir and Estefania Freites all build a character image, a clean product image and often a location image first. Eugenio Zabell's rule for the product shot is plain: straight on, with a label you can read, or the model writes its own text.
2. Let an LLM write one prompt per shot. Mike Futia, Dusan Radovanovic and Romain Torres give Claude the product, the angle and the length and get back a shot list with a prompt for each shot. Estefania Freites maps her six shots to ad beats, from a scroll-stopping close-up to a product hero for the call to action.
3. Reference every asset by name. Abhinav Kalyan uses the same labels in Claude and in the video tool, and @lr_mvcreative states what to take from each reference and what to ignore. BytePlus asks for the same thing: a clear relationship between each reference and the new video.
4. Use a reference video for motion you can't describe. @venturetwins, Fraser Cottrell and @DavidRunsAds upload a clip plus a character image and let Seedance follow its movement. Use footage you shot or have the rights to.
5. Render in batches and keep the best take. Abhinav Kalyan runs several batches per scene. For longer ads, Mike Futia starts each 15-second clip from the last frame of the previous one, and Eugenio Zabell chains clips the same way.
6. Finish in an editor. @codeMdSanto, @The_Kremlinn, Mike Futia and Estefania Freites all cut, caption and re-time in CapCut or ffmpeg rather than chasing a perfect single render.

Before launch, Nipun Taneja checks that the hook works with the sound off, then runs several variants instead of one. If you make ads for brands, the same rule applies: every ad needs a product moment and a claim the brand can legally make.

How should you write a Seedance prompt?

Write it as a short storyboard in BytePlus's order: precise subject, action, scene, lighting and colour, camera movement, visual style, image quality, then constraints. BytePlus calls a timeline-based storyboard the ideal form for a complex video, with one camera movement per shot.

Two points where many tutorials differ from the official guide:

• Timestamps. Eugenio Zabell and Estefania Freites write timed beats every few seconds. BytePlus says precise timing such as 0 to 3 seconds is unstable and suggests ordered shots with the pacing left to the model. If timed beats misbehave, switch to Shot 1, Shot 2, Shot 3.
• Character sheets. Aswad Tahir and Fraser Cottrell use multi-angle sheets. BytePlus advises a headshot plus a full-body photo instead, and warns that multi-view images can be read as different people, which makes face drift worse.

Spoken lines go in curly braces after the speaker, and the constraints at the end can ask for no subtitles, logos or watermarks. @BLOCKLORDS gets usable native audio for short videos by writing the tone and the line into the prompt. Mike Futia also bans words that make footage look staged, such as cinematic and bokeh, when he wants a handheld street look.

The full templates live on our Seedance prompts page, so we don't repeat them here. For a single shot, the AI ad prompt builder fills BytePlus's order for you. To estimate what a batch of takes will cost, use the AI video cost calculator.

Common problems and fixes

Most failures in the tutorials are consistency problems, and most fixes happen in the reference images, not the prompt.

What creators report

These are the authors' own claims, not tested by Anglr.

• Tyler Reed reports that Seedance 2 gave more natural phone-style UGC footage than Kling 3.0 from the same start image, while Kling did better on product and effects shots.
• @drchibby_ reports that Seedance 2.5 acts more realistically than Seedance 2 from text alone.
• Mike Futia reports shipping a UGC ad in about 20 minutes with his Claude Code skill.
• Estefania Freites reports making 12 multi-scene ads in a week from production boards.

Fast renders are only half the job. The ads that earn are tested scripts, cut cleanly. If your Seedance ads already perform, Anglr pays AI creators on performance to remake winning scripts: see what we look for. More workflows are on AI ad guides, including AI UGC ads and AI product videos.