Seedance for AI Micro Drama: 2.5 vs 2.0, Pricing and 9:16 Tips

Use Seedance when your scene has people talking. It makes the picture and the sound in one go, with lip sync (mouth movements that match the words). It also copies faces from reference images (pictures you give the model so it keeps the same face).

Since version 2.5, one clip can run 30 seconds. That's a third of a typical vertical episode.

Below is what it's good and bad at, what it costs, and how to set it up so episode 40 still looks like episode 1.

What it's good at, and what it isn't

Use it for:

• Talking scenes. Seedance 2.x makes voices, sound effects and music along with the picture. So the slap, the gasp and the line after it all land in sync.
• Several angles in one prompt. Write "Shot 1, Shot 2, Shot 3" and it cuts between angles with the same people in the same room. Version 2.5 also follows second-by-second timestamps.
• Lots of references. Seedance 2.0 takes up to 9 images, 3 videos and 3 audio clips per request. Seedance 2.5 takes up to 30 images, 10 videos and 10 audio clips.
• Long takes. One clip runs up to 30 seconds, and you can extend it (add more seconds to the end). AI Filmmaking Academy made a full wedding-toast episode that way in CapCut Web: a 90-second scene from one clip extended twice.

Watch out for:

• Real faces. BytePlus won't take reference images or videos with real human faces, and the CapCut version blocks them too. So you can't upload a photo of your actor (workarounds below).
• Crowds. ByteDance's own 2.0 guide says results get shaky past four referenced people. The 2.5 launch post admits scenes with several people still need work.
• Text. Signs and phone screens come out misspelled. Unwanted subtitles also show up more often in vertical video than in horizontal.
• Extending again and again. Each extension makes the image a little softer. Faces show it first.

Current versions

Use 2.5 for your main scenes and 2.0 Mini or Fast for cheap drafts.

All of these output 24 fps and support 9:16. Version 2.5 also has a draft mode. You make a cheap 480p preview first, then finish the same take at 1080p.

Where to run it

• Dreamina (CapCut's web app). The easiest way in: you buy credits, you don't need code, and CapCut's editor is right there. Prices and deals change by country and by month, so check the pricing page before you budget.
• BytePlus ModelArk. ByteDance's official API (you pay per clip from code instead of using an app), with the lowest list prices. To turn on 2.5, you need a balance over $30, a $30+ savings plan, or a resource pack.
• fal.ai. Pay per second, with US-hosted endpoints if you care where your requests run. It also offers 2.5's draft mode.
• Replicate. Official ByteDance listings, billed per second of video.
• Creator platforms like Higgsfield and OpenArt resell Seedance through their own credits. See the tools page.

Dreamina's deals can be big. A late-September 2026 offer listed 720p Seedance 2.5 at about $0.035 per generated second for the first month.

One BytePlus gotcha: your video links expire after 24 hours, so download right away.

Pricing (as of October 2026)

Here's the cost of 9:16 video at 720p with no video input. Each cell shows the price per second, then the cost of a 10-second clip.

The takeaway: BytePlus and Replicate charge about half of what fal does for the same model.

BytePlus bills in tokens (small units of video), not seconds. Tokens are seconds × width × height × 24 ÷ 1,024. A 720×1280 frame works out to 21,600 tokens per second.

So Seedance 2.5 at $10.70 per million tokens costs about $0.231 a second. At 1080p, 2.5 costs $11.70 per million, or roughly $0.57 a second. You don't pay for failed generations, including ones blocked by moderation.

BytePlus also ran discounts on 2.0 Fast and Mini for enterprise customers only, through October 7, 2026.

Plan for retakes. If you keep one take in three, a 90-second episode on 2.5 through BytePlus costs about 90 × $0.231 × 3. That's roughly $62 before music and editing.

The cost breakdown guide compares that with other models.

Best practices for 9:16 serialized drama

Build the cast once, then reuse it

• Make a clean headshot and a full-body image of each lead. Use a plain background and no text. ByteDance's guide says that pair is enough.
• On 2.0, skip turnaround sheets (one image showing the face from several angles), because they can give you twins. Version 2.5 handles them fine.
• Real faces are blocked, so design your characters with an image model. On BytePlus, faces made on the same account in the last 30 days count as trusted inputs.
• Keep those original files, because edited or recompressed copies lose that trust. BytePlus also has a library of preset digital characters you can use.
• Name every reference in the prompt: "@Image 1 is Elena's face, @Image 2 is Elena's outfit." Upload them in the order the characters first appear.
• Put signature props in the character image itself. A ring or a scar that's in the reference shows up far more reliably than one you ask for in text.
• Don't fill every reference slot. The 2.0 guide suggests four or five files per request, and creators say one to three character images work best.
• Use one model version for the whole season. Seedance 2.0 and 2.5 look visibly different, and cutting between them looks like a reshoot.

Write prompts like a shot list

ByteDance suggests this order: subject, action, scene, style, camera, sound. On 2.0, split the clip into Shot 1, Shot 2, Shot 3. On 2.5, use timestamps in whole seconds ("0–3s", "3–7s").

Don't cram too much into one time window, or the model will skip beats. Name shots in plain terms: close-up, over-the-shoulder, push in, slow orbit.

Show emotion through the body: "jaw tight, eyes wet, she doesn't blink." Big emotion words can give characters glowing eyes, which the 2.5 guide lists as a known issue.

Frame for the phone

• Set the aspect ratio to 9:16 for text-to-video and reference-to-video. For image-to-video (you give it a still and it animates it), the video copies the shape of your start frame. So make that frame 720×1280 or 1080×1920.
• Keep faces in the upper middle of the frame. Captions and app buttons cover the bottom.
• Add "no subtitles, no on-screen text" to every vertical prompt. There's more in vertical framing and shot language.

Dialogue and lip sync

Put spoken lines in curly braces and name the language, like says in English {You signed it without reading it?}. That's the official format. Keep it to one short line per beat.

Don't repeat words from the line elsewhere in the prompt. Don't attach instructions to single words either. Both make unwanted subtitles more likely.

Voices drift (sound a bit different) from clip to clip. The fix is an audio reference. Creators testing 2.0 got the closest match when the uploaded clip already had the exact line in the right voice.

If you plan to dub, write "no BGM, no music, no score" at the start and end of the prompt (BGM means background music). Then add music in the edit. Our lip-sync and dubbing guide covers the rest.

Length, extension and stitching

A 60 to 90-second episode is three to six clips on 2.0, or two or three on 2.5. Extend a take once or twice at most, then start the next shot fresh.

Joins between extended clips can jump. ByteDance's fix is to trim about 6 frames from the end of the first clip and 1 from the start of the next. Cut on a reaction, not in the middle of a gesture.

Sometimes a click lands at the very end of the audio. Fade out the last few frames of sound to get rid of it.

Pacing

Start every clip in the middle of the action, because the first three seconds decide whether people swipe away. End on a held reaction close-up. That way you always have a clean place to cut.

Our script guide and script templates are built around those beats.

Common pitfalls

1. Uploading actor photos. They're blocked. Use generated or verified characters instead.
2. Too many people. Past four, you'll get twins or a missing extra. Add a headcount line like "exactly two people in frame."
3. Story text in the frame. Contracts, texts and headlines come out garbled. Add them as overlays in the edit.
4. Mixed languages. Keep each clip's dialogue in one language.
5. Water and echo cues. The 2.5 guide links them to bubbling sounds in the audio. Add the rain in your sound mix instead.
6. Huge reference images. Very high-resolution AI images can leave fingerprint-like textures. Shrink references to your output resolution.

Prompts and next steps

Copy-ready prompts are in the Seedance prompt library. They're sorted by genre: CEO romance, revenge, werewolf, historical, thriller, comedy, fantasy and family.

For the bigger picture, read character consistency and choosing a video model. You can also compare Seedance with Kling and Veo.