Why 30 seconds need a different prompt
A model making 5 to 10 seconds needs one action. Seedance 2.5 makes up to 30 seconds in a single pass, and the developer positions this as a way to tell a short story: set-up, development, ending. If the prompt describes one moment, the model stretches it over half a minute. If it describes ten events, the model crams them and skips some.
The solution: split the clip into time segments. ByteDance does exactly this in its official examples for the model.
Prompt structure
[General line: format, style, type of shot]
0-8s: [where we are, who is in frame, where the camera is, what is happening]
8-18s: [one action, one camera move]
18-26s: [development or a turn]
26-30s: [final shot]
[Sound: background, music, lines]
Rules:
- One action and one camera move per segment.
- The general line sets everything that does not change: aspect ratio, style, type of shot (one continuous take or edited).
- The final shot separately. The model ends the clip better when it knows where to stop.
- Sound at the end. If you do not describe it, the model will invent it.
General rules for video prompts: how to write prompts for video.
References
Seedance 2.5 accepts many references: according to the developer, up to 30 images, 10 videos and 10 audio files. The set on a particular platform may be smaller.
- One role per reference: heroine, location, style, object.
- Refer to them by number: «the woman from image 1», «the kitchen from image 2».
- For several heroes, give each their own reference and say who stands where.
How to keep a character consistent: character consistency.
10 ready prompts
The prompts are in English; plug in your own details. All the heroes are fictional.
1. A product ad in one take
16:9, photorealistic commercial, single continuous take, smooth gimbal movement.
0-8s: Close-up of a steel thermo mug on a wooden kitchen counter at dawn, steam rising, soft window light. Slow dolly in.
8-18s: A woman's hand picks up the mug, camera follows the mug as she walks to the hallway and puts it into a backpack.
18-26s: Camera tracks her stepping outside into a snowy street, she takes a sip, breath visible in cold air.
26-30s: Slow push-in on the mug in her hand, warm light, clean space on the right for text.
Sound: quiet morning ambience, soft acoustic guitar, footsteps on snow.
Why this works: the product is in every segment, and the «morning, commute, cold» story shows the benefit without words. More on ads: AI product ad video.
2. Moving through locations
9:16, cinematic realism, one continuous handheld take, no cuts.
0-8s: A young barista with curly hair opens the door of a small coffee shop at sunrise, camera follows from behind.
8-18s: She walks behind the counter, turns on the espresso machine, camera circles around her to the front.
18-26s: She hands a cup to the first customer, smiles, camera pulls back to show the whole cozy room.
26-30s: Wide shot through the window from the street, warm light inside, first passersby outside.
Sound: door bell, espresso machine hiss, soft café music.
3. A character from a reference
16:9, cinematic, soft natural light. The man from image 1 in the cabin from image 2.
0-8s: Medium shot, he sits by the window reading a letter, rain outside. Static camera.
8-18s: He stands up, walks to the old radio, turns it on. Slow tracking shot.
18-26s: Close-up of his face as music starts, a slight smile appears.
26-30s: Wide shot of the cabin from outside in the rain, warm window glowing.
Sound: rain on the roof, old radio crackle, gentle piano.
4. A cartoon for children
16:9, soft 3D animation, bright pastel colors, gentle light.
0-8s: A small round robot wakes up in a toy box, eyes light up blue. Camera slowly pushes in.
8-18s: The robot climbs out and rolls across the bedroom floor, chasing a rolling ball.
18-26s: The ball stops at the feet of a sleepy orange kitten, the robot and the kitten look at each other.
26-30s: The kitten taps the ball back, the robot spins with joy. Camera pulls out to a wide shot.
Sound: playful xylophone melody, soft beeps, a tiny meow.
On cartoons as a whole: how to make an animated film with AI.
5. Food for a restaurant
9:16, food commercial, macro details, warm restaurant light.
0-8s: Extreme close-up of fresh pasta dough being cut by a chef's knife. Slow motion.
8-18s: Pasta drops into boiling water, camera follows the steam up to the chef's hands adding sauce in a pan.
18-26s: Plating: the chef twirls pasta onto a white plate, sprinkles parmesan. Top-down shot.
26-30s: Slow push-in on the finished plate on a wooden table, candlelight in the background.
Sound: knife on board, bubbling water, sizzling pan, soft jazz.
6. Real estate
16:9, real estate tour, bright daylight, smooth stabilized camera, single continuous take.
0-8s: Camera moves from the front door into a bright hallway with light wooden floor.
8-18s: Glides into a spacious living room with large windows, sunlight on a beige sofa.
18-26s: Turns into an open kitchen with white cabinets, slows near the island.
26-30s: Moves to the balcony door, which opens to a city view at golden hour.
Sound: calm ambient music, distant city sounds on the balcony.
7. A music video
9:16, music video, neon night city, rain, handheld energy.
0-8s: A girl in a yellow raincoat walks under neon signs, camera tracks backwards in front of her.
8-18s: She starts dancing on the wet street, reflections in puddles, camera circles around her.
18-26s: Close-ups intercut: her shoes splashing water, her hands, her face looking up at the rain.
26-30s: She stops, looks straight into the camera, neon lights flicker off.
Sound: synthwave beat, rain, splashing steps.
8. Nature and atmosphere
16:9, nature documentary, misty morning.
0-8s: Aerial shot over a foggy pine forest at sunrise, slow forward movement.
8-18s: Camera descends to a calm lake, mist on the water surface.
18-26s: A deer walks to the shore and drinks, camera holds at a distance.
26-30s: The deer lifts its head, sun breaks through the mist. Slow pull back.
Sound: birdsong, quiet water, soft wind.
9. Extending a clip
If the platform supports extension, describe the next segment as a continuation:
Continue the previous video. Keep the same characters, location, lighting and sound style.
0-10s: The robot and the kitten roll the ball into the hallway together.
10-22s: They find a cardboard box and climb inside, the box starts moving.
22-30s: The box stops in the sunlit kitchen, both peek out. Wide shot.
10. A short 6-second draft
Before an expensive 30-second clip, test the key scene briefly and cheaply:
16:9, photorealistic, soft window light.
A steel thermo mug on a wooden counter, steam rising, slow dolly in. Quiet morning ambience.
6 seconds at 480p on Seedance 2.5 costs 102 tokens, about 161 rubles. If the scene works, move it into the long prompt.
Typical mistakes
- Too many events. The model crams more than five segments in 30 seconds.
- Empty words. «Epic, cinematic masterpiece, 8K» give nothing specific.
- Contradictions. «Static camera» in the general line and «camera circles» in a segment.
- Real people and brands. Recognisable faces, logos and famous characters often lead to refusals. On KADR Seedance 2.5 runs with a soft filter that refuses ordinary people less often, but that does not apply to other people’s celebrities and brands. More: other brands, logos and characters in AI.
- No final shot. The clip breaks off in the middle of an action.
What attempts cost
Seedance 2.5 prices on KADR, October 2026, sound included in the price:
| Resolution | 6 seconds | 30 seconds |
|---|---|---|
| 480p | ~161 rub. | ~794 rub. |
| 720p | ~341 rub. | ~1700 rub. |
| 1080p | ~849 rub. | ~4242 rub. |
This is not a public offer.
The order that saves money: a short draft of the key scene at 480p, then the full 30-second prompt at 480p, and only then the final at 1080p. Model review: Seedance 2.5. Comparison with other models over 30 seconds: Seedance 2.5, Kling or Veo.
In short
A Seedance 2.5 prompt is written by timecodes: a general line, 3 or 4 segments with one action and one camera move each, a final shot and a description of the sound. Number your references and give each one role. Test the scene briefly at 480p first, and run the long clip at 1080p once the prompt already works.
