Seedance 2.5 Prompt Guide: Structure, Examples and Real Clips

Write a Seedance 2.5 prompt in ByteDance's order: subject, action, scene, style, camera, sound. Mark each shot with a timestamp or a "Shot N" line, and put spoken lines in double quotes. We ran six prompts at 720P on October 26, 2026: the timed cut, the spoken line and the sign text all came out as written.

The Seedance 2.5 prompt structure

ByteDance's tutorial orders a prompt as subject, action or event, scene, visual style, camera and shot cuts, then sound, and unneeded parts can be left out. Its prompt guide shortens that to subject + location + event + genre/style + camera movement. What else ByteDance confirmed for 2.5, such as 30-second clips and up to 50 reference files, is in our Seedance 2.5 review.

• Time a clip with whole-second timestamps that leave no gaps, such as "0s-2s" then "2s-5s", or with "Shot N" lines. Points ("at the 2-second mark") and relative times ("after 3 seconds") work too. Do not time fast repeated actions, such as a head shaking three times a second.
• Seedance 2.0 ignores timestamps and follows only shot numbers, so use "Shot N" if a prompt has to work on both.
• Describe what you want to see. ByteDance supports negative constraints only for subtitles and audio, such as "No subtitles" and "No BGM".
• Put spoken lines in double quotes. We put the sign text in quotes too, and it came out exactly.
• ByteDance recommends at most 500 Chinese characters or 1,000 English words, because long text scatters detail. No hard limit is published.
• The tutorial lists 11 prompt languages: Chinese, English, Spanish, Indonesian, Malay, Thai, Arabic, Portuguese, Vietnamese, Japanese and Korean.

The tutorial also marks audio layers with brackets: () for music, <> for sound effects, {} for dialogue and 【】 for subtitles. We did not test the brackets.

How we made the clips

• Model dreamina-seedance-2-5-260628 through OmniaKey's async media API on October 26, 2026 (Beijing time), with duration 5, resolution 720P and generate_audio true for every clip. Clips 1–4 and 6 ran text_to_video at 16:9; clip 5 ran image_to_video at adaptive with a Nano Banana 2.1 still as the first frame.
• Every task succeeded on the first try: MP4, H.264, 24 fps, 5.04 s, mono AAC. Clips 1–4 and 6 are 1280×720; clip 5 came back at 1286×716, the still's shape.
• Generation took 153.6 s to 252.0 s per clip (clip 1 154.5 s, 2 153.6 s, 3 246.3 s, 4 252.0 s, 5 155.6 s, 6 198.0 s).
• We pulled frames at fixed times and viewed them at full size, ran a scene-cut detector on every clip, and checked each soundtrack with two Gemini audio models and a spectrogram. Every sound claim below is one those checks agreed on.
• OmniaKey sells API access to Seedance 2.5, so we ran the clips through it. The prices below are BytePlus's list prices, not ours.
• Sources: ByteDance's Seedance 2.5 prompt guide, the Seedance 2.5 tutorial, the create-video-task API reference and the model pricing page, checked October 25, 2026.

We ran each prompt once, so each result is one sample.

Six prompts with real clips

1. One sentence: subject, place, action, style, camera

Clip 1 puts the whole formula in one sentence.

A golden retriever puppy chases a red ball across a sunlit backyard lawn, warm home-video style, low tracking shot that follows the puppy.

Settings: text_to_video, 16:9, 720P, 5 s, audio on.

Result: the subject and the place match; the look and the camera move changed.

• The puppy, the red ball and the sunlit lawn are there, plus a wooden fence and a garden chair, filmed from a low camera. The look is polished golden-hour footage, not home video.
• The puppy trots toward the camera behind the rolling ball, and the camera moves back ahead of it.
• Sound: a cheerful instrumental music bed that the prompt never asked for.

What to change:

• Add a sound line with "No BGM". The four clips below that said it have only effects and speech.
• To follow from behind, try "the camera follows behind the puppy at ground level".

2. Timestamps inside one shot

Clip 2 times two beats inside one continuous shot.

Commercial product video, warm studio lighting. A barista's hand pours steamed milk into a white ceramic cup of espresso on a dark walnut counter.

0s-2s: Side close-up. The milk stream falls into the espresso and a white heart shape starts to form on the surface.

2s-5s: The pour ends. The camera slowly pushes in from above as the finished latte art settles and a thin wisp of steam rises.

Sound: soft pouring and a quiet café murmur. No BGM. No subtitles.

Settings: text_to_video, 16:9, 720P, 5 s, audio on.

Result: both beats happened, about a second late and from a higher angle.

• A hand pours milk from a steel jug into a white cup on dark wood, and the swirl becomes a heart.
• The view is from high above at an angle, where the prompt asked for a side close-up. The pour ends at about 3 s, a second after the 2 s mark, and then the camera looks straight down on the finished heart.
• The frames show no steam. Sound: pouring, no music.

What to change:

• Give the angle its own sentence in each segment: "Camera at cup height, side view."
• Leave a second of slack, because "0s-2s" ran to about three seconds here.

3. A named camera move: orbit

Clip 3 gives the camera one named move, a 180-degree orbit.

A street dancer in a white hoodie does a fast spin and a freeze on the top floor of an empty concrete parking garage at sunset. During the move, the camera orbits 180 degrees around the dancer at chest height, keeping the dancer centered. Warm orange backlight, long shadows, cinematic look. Sound: sneakers squeaking on concrete and distant traffic.

Settings: text_to_video, 16:9, 720P, 5 s, audio on.

Result: the orbit and the light match; the distance and the location changed.

• A dancer in a cream-white hoodie spins and ends in a low freeze with one hand on the ground, silhouetted against the setting sun with long shadows.
• The garage has a concrete ceiling: a covered level, where the prompt asked for the open top floor.
• The camera moves around the dancer, so the background swings from the garage wall to the open side with the sunset and the city. It also pulls back to a full-body wide shot, so the dancer changes size.
• Sound: no music. The two audio models heard different effects, so the squeaks and the traffic stay unconfirmed.

What to change:

• To keep the distance fixed, try "the camera keeps the same distance as it orbits".
• For an open roof, try "open sky above, no ceiling".

4. A spoken line and sign text in quotes

Clip 4 tests a spoken line and a price in quotes.

A small neighborhood bakery at 7 a.m. Medium close-up of a smiling baker in a flour-dusted apron behind the counter. A chalkboard behind the counter reads "SOURDOUGH $6" in white chalk. The baker looks into the camera and says: "Fresh sourdough just came out of the oven." Then the baker lifts a round loaf toward the camera. Warm window light, realistic. Sound: the baker's voice, the crackle of the crust and quiet street noise outside. No BGM. No subtitles.

Settings: text_to_video, 16:9, 720P, 5 s, audio on.

Result: the line, the sign and the order of actions all match the prompt.

• A smiling man in a white T-shirt and a beige apron stands behind a counter, and the chalkboard reads exactly "SOURDOUGH $6".
• He says "Fresh sourdough just came out of the oven." with his mouth moving, then holds a large scored loaf out toward the camera, in the order the prompt gave.
• The flour on the apron is missing. Sound: the line, with no music and no subtitles on screen.

What to change:

• For longer or unusual words on a sign, ByteDance's guide says to spell them letter by letter or to give the text as an image.
• Keep to one spoken line per few seconds of clip.

5. Image-to-video: let the still set the look

Clip 5 starts from a still, so the prompt only has to carry motion, camera and sound. The first frame is a still we made with Nano Banana 2.1 at 16:9 and 2K from the prompt "A white paper airplane resting on a wooden desk next to an open window, soft morning light, a city skyline softly blurred outside, realistic photo."

The paper airplane lifts off the desk, glides out through the open window and drifts over the city. The camera follows it to the window and stops at the sill. Sound: a gentle breeze and distant birds. No BGM.

Settings: image_to_video, adaptive, 720P, 5 s, audio on.

Result: the still set the look and the prompt set the motion.

• The clip opens on the still's exact framing and keeps its look. Every step of the prompt happens, in order: the plane lifts off, glides out of the window and shrinks into the sky over the rooftops.
• The camera moves up to the window and stops with the sill at the bottom of the frame.
• Sound: no music. The spectrogram shows short repeated chirps after the first second.

What to keep:

• Describe only motion, camera and sound, because the still already fixes the look.
• Make the still in the shape you want the video, because adaptive keeps the still's shape.

6. Two shots and a timed cut

Clip 6 is the multi-shot test: two "Shot N" lines and a cut at 2 s.

Cinematic short film, a lighthouse on a rocky coast at dusk.

Shot 1 (0s-2s): Extreme wide shot from the sea. Waves crash against the rocks and the lighthouse stands dark against an orange sky.

Shot 2 (2s-5s): Cut to a close-up inside the lantern room. A keeper's weathered hands turn a brass switch, and the big glass lens begins to glow and rotate, casting a beam across the window.

Sound: waves and wind, then a low mechanical hum when the lamp starts. No BGM. No subtitles.

Settings: text_to_video, 16:9, 720P, 5 s, audio on.

Result: both shots and the cut came out, with one continuity slip.

• The cut detector found the cut at 2.29 s. Shot 1 is a wide view from the sea at dusk: dark cliffs, big waves breaking, the lighthouse a dark silhouette against a pink-orange sky.
• Shot 2 is a close-up of a weathered hand on a brass handle in front of a large ribbed glass lens, which is dark at first and glows warm and bright by the end.
• Sound: no music. The steady noise of shot 1 stops at the cut, and shot 2 has a click and a rising hum.
• One continuity slip: the lamp at the top of the lighthouse is already lit in shot 1.

What to change:

• Say the state you want in shot 1 as a positive: "the lamp at the top is dark".

Common failures and fixes

ByteDance's prompt guide ends with a list of known problems and workarounds. These are the ones a prompt can fix, plus one from our clips.

What a clip costs

BytePlus bills Seedance 2.5 by the token. A clip's tokens are (input video duration + output duration) × width × height × frame rate / 1024, and the model pricing page charges $10.70 per million tokens at 480p and 720p and $11.70 at 1080p.

These are BytePlus's own worked examples, checked October 25, 2026. At that list price our six 720P clips would come to $6.94 (6 × $1.156), and BytePlus charges only for successful generations. For how Seedance 2.5 compares with Kling 3.0, see Kling 3.0 vs Seedance 2.5.

For OmniaKey's current rates, see the Seedance 2.5 model page.

Run these prompts through the API

This request sends clip 2's latte prompt. The JSON sits in a quoted heredoc, so the apostrophe in "barista's" goes in as is:

curl https://api.omniakey.com/v1/media/tasks \
  -H "Authorization: Bearer $OMNIAKEY_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: latte-pour-001" \
  --data-binary @- <<'JSON'
{
  "model": "dreamina-seedance-2-5-260628",
  "type": "video",
  "operation": "text_to_video",
  "input": {"prompt": "Commercial product video, warm studio lighting. A barista's hand pours steamed milk into a white ceramic cup of espresso on a dark walnut counter.\n\n0s-2s: Side close-up. The milk stream falls into the espresso and a white heart shape starts to form on the surface.\n\n2s-5s: The pour ends. The camera slowly pushes in from above as the finished latte art settles and a thin wisp of steam rises.\n\nSound: soft pouring and a quiet café murmur. No BGM. No subtitles."},
  "parameters": {"duration": 5, "resolution": "720P", "ratio": "16:9", "generate_audio": true}
}
JSON

1. The reply is 202 with a task id.
2. Fetch https://api.omniakey.com/v1/media/tasks/{id} with that same id until status is succeeded. Our clips took 153.6 s to 252.0 s, so keep polling instead of resubmitting.
3. Download the clip from assets[0].url.

For image-to-video, set operation to image_to_video and ratio to adaptive, and send the still in input.media with type set to first_frame and a public HTTPS URL or a data URI. The whole request body is capped at 1 MiB. Use the exact model id, because the short aliases return 503 no_available_channel, and send generate_audio: false for a silent clip. Start from the Seedance 2.5 model page, and see the video API docs for every parameter.

Common questions

How long can a Seedance 2.5 prompt be?

ByteDance recommends at most 500 Chinese characters or 1,000 English words, because in long text the model may drop details. No hard cap appears in the API reference or the prompt guide.

Can Seedance 2.5 make a 30-second video from one prompt?

Yes. The API takes duration from 4 to 30 seconds, or -1 to let the model choose, and ByteDance raised the maximum for one generation from 15 to 30 seconds. The 15-second figure belongs to Seedance 2.0.

Does Seedance 2.5 add music on its own?

It can. generate_audio defaults to true, and the model then adds voice, effects and music to match the prompt and the picture. Our clip 1 said nothing about sound and came back with a cheerful instrumental bed, so write "No BGM" to keep music out.

Does Seedance 2.5 support negative prompts?

There is no separate negative-prompt field. ByteDance asks for positive descriptions and supports negative constraints only for subtitles and audio, such as "No subtitles" and "No BGM".

Do these prompts work on Seedance 2.0?

With one rewrite: Seedance 2.0 ignores timestamps and responds only to shot numbers, so "Shot 1" carries over and "0s-2s" does not. Expect a different look as well, because ByteDance says 2.5 and 2.0 have significantly different aesthetic style preferences.

Update log

• October 26, 2026: first version, six prompts run the same day.