AI Video Playbook

Seedance

ByteDance's Seedance 2.5 and 2.0, what they are good at, the official prompt syntax, and how to spend less on them.

Seedance is the model most creators compare everything else against. Use Seedance 2.5 when a shot needs many references (a character, a product, a voice, a camera move) or a long single take up to 30 seconds. Draft on Seedance 2.0 Mini or at 480p first, because 2.5 is slow and one of the more expensive models per second.

At a glance

Prices checked on 15 Oct 2026.

ItemDetail
MakerByteDance
Current versionsSeedance 2.5 (public since 31 Jul 2026), Seedance 2.0, 2.0 Fast, 2.0 Mini. Older 1.5 Pro and 1.0 are still on the API.
Max duration2.5: 4 to 30 s per clip. 2.0, Fast, Mini: 4 to 15 s.
Resolution2.5: 480p, 720p, 1080p (10 bit). 2.0: up to 4K. Fast and Mini: 480p and 720p only.
AudioNative audio. 2.5 speaks 11 languages, including Vietnamese.
References2.5: up to 50 files (30 images, 10 videos, 10 audio). 2.0: up to 15 (9 images, 3 videos, 3 audio).
AppDreamina, CapCut (paid subscribers over 16 in Europe, Asia, the Middle East and South America; the US is not listed). Also resold by Higgsfield, Runway, Luma, Krea and others.
APIBytePlus ModelArk. 5 s at 720p: 2.5 about 1.16 USD, 2.0 about 0.76 USD, Fast 0.60 USD, Mini 0.38 USD.
Cheapest way inDreamina free tier: 120 credits per day. In Vietnam, Dreamina Basic is 187,200 VND for the first month, then 390,000 VND/month for 4,161 credits (seen on the checkout page).

Dreamina does not publish how many credits one Seedance clip costs. The price shows in the app before you generate. Check it there.

Best for

From community tests, mostly organic posts:

  • Complex scenes and "taste". Several side by side tests picked Seedance over Grok Imagine, MiniMax H3, Gemini Omni Flash and Wan3.0, especially for rain, reflections and deliberate camera moves.
  • Reference heavy shots. Assigning a job to each reference (face, location, rhythm) is where it beats other models.
  • Music driven montage. Audio references set the rhythm of cuts.
  • Following a camera you blocked in 3D. Creators who block the camera in Blender report Seedance follows it almost one to one.

Weak at

Also from community reports:

  • Speed and cost. In one organic test, 2.5 took 3 min 15 s and 22.5 credits for a shot Wan3.0 made in 58 s for 7.5 credits. Another tester found it about 4 times more expensive per second than Kling 4.0 Flash.
  • Audio control. It invents sound or ignores the audio reference in roughly 15 to 20% of runs (one organic tester).
  • Strict moderation. Prompts and references get rejected across platforms, including blood and realistic faces.
  • 2.5 is not always better than 2.0. Some testers saw worse motion and prompt following at launch, and one same prompt test preferred 2.0.
  • Fast motion. Hands and blur break in fast action.

How to prompt it

Structure

ByteDance's official formula for 2.5: subject, action, scene, visual style, camera movement or shot cuts, then sound. Keep one camera move per shot. Mixing push, pull and pan in one shot makes the image unstable.

References

Upload files, then call them by type and upload order: @Image 1, @Video 1, @Audio 1. Say what each one provides and what not to copy, for example "@Image 1 is the face only, not the clothes". Do not rely on text written inside an image.

3D clay reference

Seedance 2.5 officially accepts a "clay" video, rendered from a simple 3D scene, as a reference for camera, blocking and motion. Use plain shapes, give each stand in its own colour and map each colour to an image reference in the prompt, render without overlays, and say which parts of the clay video to use (camera only, camera and motion, or also lighting). Reference videos need at least 407,696 pixels per frame (854x480 passes, 640x360 fails) and 24 to 60 fps. See Previz in Blender.

Sound and dialogue (2.5)

You wantWrite it insideExample
Music( )(slow lo-fi piano)
Sound effect< ><rain on a tin roof>
Dialogue{ }{Cà phê của anh đây.}
Subtitles【 】【Day 1】

Name the language before any dialogue that is not Chinese, for example "she says in Vietnamese". To keep audio or text out, say it plainly: "No subtitles", "No BGM".

Timing

  • Seedance 2.5 follows whole second timestamps: "0 to 3 seconds: ...", "[1s to 4s]", "at the 5 second mark".
  • Seedance 2.0 ignores timestamps. Use "Shot 1, Shot 2" instead.

Things that do not work

  • There is no negative prompt field. Describe what you want instead.
  • "Do not generate a logo or watermark" reduces logos but does not stop them.
  • When you edit or extend a clip, call it "Video 1", not "reference Video 1", or Seedance treats it as a reference.

Example prompts

A product shot with a reference image and sound tags:

Prompt
@Image 1 is the product: keep the bottle shape, cap and label layout exactly.
A glass bottle of cold brew stands on wet black slate. Water drops slide down the glass,
a slow push in ends on the label. Single hard rim light from behind, dark studio.
<ice cubes clinking> <a drop hitting the slate>
No subtitles. No BGM. Do not generate a logo or watermark.

A three shot sequence with timestamps (2.5 only):

Prompt
@Image 1 is the identity of the man only. @Image 2 is the location and lighting only.
0 to 4 seconds: wide, static. The man walks onto a rain wet ferry pier at night and stops at the railing.
4 to 8 seconds: medium close up, 85mm, slow push in. He breathes out, his breath visible, and looks left.
8 to 12 seconds: over the shoulder. Ferry lights sweep across his face.
<drizzle> <distant ferry horn> <footsteps on wet wood>
No BGM. No subtitles.

Vietnamese dialogue in a café:

Prompt
@Image 1 is the barista: keep her face, hair and mustard apron.
Medium shot, static camera. In a small Saigon café in the morning, the barista slides a glass
of iced milk coffee across the counter and smiles.
She says in Vietnamese: {Cà phê sữa đá của anh đây.}
<ice clinking> <street traffic outside, soft>
(quiet acoustic guitar, low in the mix)
No subtitles.

The same idea for Seedance 2.0, which needs shots instead of timestamps:

Prompt
@Image 1 is the identity of the girl. @Audio 1 sets the rhythm of the cuts only.
Shot 1: wide, a girl in a red raincoat runs across a flooded street, splashing.
Shot 2: close up, she laughs and wipes rain from her face.
Shot 3: low angle, she jumps into a puddle, water sprays toward the camera.
Hand held camera, overcast daylight, muted colors. No subtitles.

Settings that save credits

  • Draft at 480p. On the API, a 5 s clip of 2.5 costs about 0.51 USD at 480p and 2.84 USD at 1080p. Pick the take at 480p, then render the final once.
  • Use Draft mode on 2.5. It makes a 480p preview, then the final. You pay for both, but the draft is billed at 480p.
  • Rough out on 2.0 Mini. Mini is the cheapest Seedance (0.38 USD for 5 s at 720p on the API). One creator roughs out on Mini, then moves the keeper to 2.5.
  • Keep clips short while testing. 5 s is enough to judge style and motion.
  • Audio does not change the API price on Seedance 2.x, so leave it on if you need it.
  • Failed and moderated generations are not charged on the API.

Known issues

ProblemWhat to do
Subtitles appear on their ownAdd "No subtitles"
Background music despite "No BGM"Add the sounds you do want with < >, then mute or replace music in the edit
The wrong character takes a reference's faceNumber the references in upload order and give each one a single job
Misspelled text on signs and packagingAdd text in the edit instead
Glowing eyesA known issue in ByteDance's own FAQ. Regenerate, and check the take before you upscale it.
Mixed languages in dialogueName the language before each line
Real faces rejected2.5 does not accept uploaded images or videos of real people. Use generated or stylized characters.
Prompt rejected several timesRemove violence, blood and real person references, then retry

API outputs are kept for 24 hours. Download them right away.

Sources