Seedance 2.5 prompts: 6 tests, 10 prompts, and how to write your own
Seedance 2.5 is now on Magnific. Thirty seconds of uninterrupted video from a single generation. Characters that stay themselves from the first frame to the last. On-screen text that holds, letter by letter, through motion and cuts.
We didn’t take ByteDance’s word for it. We ran six tests designed to find the edges: a movie trailer, a product ad, two mini films, a physics stress test, a character interaction, and a music video that had no business working. Every prompt is in this post, written to ByteDance’s official Seedance 2.5 prompt spec, ready to copy straight into Magnific.
Came here from the video? Jump straight to the prompts.
Table of contents
- What Seedance 2.5 changes
- Where it still needs you
- The prompts
- How to write Seedance 2.5 prompts
- Seedance 2.5 FAQ
What Seedance 2.5 changes
Three things stood out across every test.
One generation, one story
The F1 trailer is 30 seconds of continuous video from a single prompt. Not stitched, not extended. One pass. The model built the narrative arc, held the characters consistent through every scene, and cut between shots on its own. Nobody told it to.
Text that behaves
This was the weak point of every video model before it. In the shaved-horse test, “Seedance 2.5” appears carved into the horse’s coat, stars included, every star where it should be, and stays legible while the horse moves. In the sunglasses ad, spec callouts appear as thin white lines pointing at the frame, in slow motion, and hold. The rapper knight’s silver grillz survive every cut.
One reference goes far
The sunglasses commercial came from a single image of a woman wearing glasses. The sneaker ad kept the shoes consistent through an entire parkour run off one reference. More references improve product consistency, but the floor is already high.
Where it still needs you
Seedance 2.5 is unusually literal. It does what you ask, which means most bad output is an underspecified prompt, not a model failure. Two things to know before you generate.
High-action scenes want smaller pieces. Our falling-car test packed too much into one shot: door panels morphed slightly, and an alligator made a manhole cover lift on its own. Complex action holds better broken into stages, each with one primary event.
Don’t give text creative freedom. If words appear on screen, write them out exactly. The model renders what you specify with precision, and improvises when you don’t.
The prompts
Every prompt below follows BytePlus official spec for Seedance 2.5: one primary event per stage, a visible end state after each one, a consistency block, and the model’s audio syntax. Music goes in round brackets, sound effects in angle brackets, dialogue in braces, and on-screen text in 【 】 brackets. One more rule before you run them: duration and aspect ratio are set on the generation page, never in the prompt. Set 30 seconds there.
Copy them, run them, break them, rewrite them. All of them run on Magnific’s AI video generator with Seedance 2.5 selected.
Chaos city
VIDEO PROMPT
[Generation Goal]
Generate a fast-paced montage of wild animals appearing in unexpected places across a modern city.
[Stage 1]
Initial state: a busy downtown street, traffic moving.
Primary event: a lion steps onto the crossing and a car stops suddenly in front of it.
End state: the car is stationary; the lion stands facing it.
[Stage 2]
Primary event: an alligator crawls out of an open sewer onto the asphalt.
End state: the alligator is fully out on the street, the manhole cover resting beside the opening.
[Stage 3]
Primary event: an elephant eats leaves from a tree in Central Park.
End state: the elephant stands under the tree, trunk raised into the branches.
[Stage 4]
Primary event: a huge giraffe runs down the middle of a city street.
End state: the giraffe passes the camera, buildings rising on both sides.
[Stage 5]
Primary event: aerial view of a whale swimming in the river that crosses the city.
End state: the whale’s back breaks the surface in the middle of the river.
[Stage 6]
Primary event: penguins waddle between the tables inside an ice cream shop.
End state: the penguins gather at the counter.
[Maintain Consistency]
Keep the same city, the same daylight and a realistic photographic style across all stages. One animal species per stage.
Urban daytime light, realistic textures, documentary feel. Cut between wide establishing shots and street-level close-ups.
<Car brakes screech> <City traffic ambience> (Playful orchestral music builds across the montage)
Sneakers
VIDEO PROMPT — 1 REFERENCE IMAGE
@Image 1 defines the sneakers’ structure, materials and colours. Do not use the image background. Do not use any person in the image.
[Generation Goal]
Generate a sneaker commercial. A young athlete wearing the sneakers from @Image 1 runs parkour through a city.
[Stage 1]
Initial state: the athlete crouches at a rooftop edge at golden hour, the sneakers in frame.
Primary event: he launches into a run and vaults across a gap between two buildings.
End state: he lands on the next rooftop, both sneakers visible on the landing.
[Stage 2]
Continue from the previous stage: same athlete, same outfit, same sneakers.
Primary event: he runs several steps along a skyscraper’s glass wall.
End state: he pushes off the wall and lands back on the pavement.
[Stage 3]
Primary event: he sprints toward the camera and slides to a stop.
End state: close-up on the sneakers, laces settling, the city out of focus behind them.
[Maintain Consistency]
Keep the sneakers identical to @Image 1 in every shot: same structure, materials and colours. Same athlete and clothing throughout.
Dynamic tracking shots, low angles, speed ramps on the landings.
<Footsteps hit concrete> <Wind rushes past> (Driving electronic beat)
Sci-fi trailer
VIDEO PROMPT
An American sci-fi movie trailer about aliens arriving on Earth. Alien ships descend over a city skyline, fighter jets scramble, crowds run through the streets, and soldiers regroup at a night barricade.
High-contrast blockbuster grade, volumetric light beams from the ships, practical smoke and debris in the streets.
Quick cuts between aerial shots of the ships, ground-level crowd panic and close-ups on the soldiers’ faces.
(Tense rising trailer score with percussive hits) <A deep alien horn resonates across the city> <Jet engines roar overhead>
Rapper knight
VIDEO PROMPT
[Generation Goal]
Generate an American movie trailer about an evil god disguised as a medieval knight who wants to conquer the world. He wears silver grillz, carries a big silver sword and rides a black horse with silver armour and white glowing eyes. The movie is called “The last battle”.
[Stage 1]
Initial state: a battlefield at dusk, banners moving in the wind.
Primary event: the knight rides the armoured black horse across the battlefield.
End state: the knight halts on a ridge, sword raised, the horse’s white eyes glowing.
[Stage 2]
Continue from the previous stage: same knight, same armour, same horse.
Primary event: a slow-motion fight scene; the knight cuts through opponents, his silver grillz catching the light when he grins.
End state: the knight stands alone among fallen enemies.
[Stage 3]
Primary event: quick cuts of armies charging, the horse rearing and the knight’s face in close-up.
Dialogue language: American English. A deep voice-over says slowly: {The last battle… is now}
End state: black frame. The movie title appears in sparkling medieval typography: 【The Last Battle】
[Maintain Consistency]
Keep the knight’s face, silver grillz and silver sword, and the horse’s black coat, silver armour and white glowing eyes identical in every shot.
Epic cinematic grade, slow motion on the fights, camera angles ranging from aerial to close-up.
(Dark orchestral trailer score with heavy drums) <Swords clash> <The horse neighs>
Office trailer
VIDEO PROMPT
[Generation Goal]
Generate a comedic movie trailer about a chaotic office where everyone is waiting for the holidays, but aliens invade Earth the moment the holidays arrive.
[Stage 1]
Initial state: a crowded open-plan office, desks piled with paper.
Primary event: stressed workers rush between desks with phones ringing, while a wall calendar counts down to the holidays.
End state: a worker crosses out the last day on the calendar.
[Stage 2]
Continue from the previous stage: same office, same group of workers.
Primary event: the workers celebrate, grabbing bags and sunglasses to leave.
End state: the whole team stands at the exit, the door half open.
[Stage 3]
Primary event: alien ships appear outside the office windows and chaos erupts, papers flying, workers diving under desks.
End state: in the middle of the chaos, a giant inflatable text spelling “BUSY” bounces up like a bubble in the centre of the office.
[Maintain Consistency]
Keep the same office space, the same group of workers and their clothing across all stages. Render the inflatable text exactly as written: “BUSY”.
Bright comedy grade, handheld camera, fast cuts.
(Upbeat quirky trailer music that turns tense when the aliens appear) <An alarm blares> <Glass rattles>
F1 trailer
VIDEO PROMPT
[Generation Goal]
Generate an American movie trailer about a Formula-style race between two rival drivers. The movie is called “The final race”.
[Stage 1]
Initial state: the pit lane before the race, mechanics working around two race cars.
Primary event: the two drivers face each other in the paddock.
Dialogue language: American English. The first driver says with quiet intensity: {This is the last time you see my rear wing.}
End state: both drivers lower their visors.
[Stage 2]
Continue from the previous stage: same drivers, same helmet designs, same car liveries.
Primary event: the race starts; the cars launch off the grid and fight through the first corners.
End state: the two cars run side by side down a straight.
[Stage 3]
Primary event: quick cuts between cockpit close-ups, crowd reactions and an overtake on the final lap.
End state: black frame. The movie title appears on screen: 【The Final Race】
[Maintain Consistency]
Keep both drivers’ faces, helmet designs and car liveries identical in every shot.
Blockbuster grade, low tracking shots at track level, cockpit POV, aerial views of the circuit.
(Epic trailer score building to the finish) <Engines scream past> <The crowd roars>
Sunglasses commercial
VIDEO PROMPT — 1 REFERENCE IMAGE
@Image 1 defines the model’s appearance and the sunglasses’ frame structure, materials and colours. Do not use the image background.
[Generation Goal]
Generate a sunglasses commercial built around the model and the sunglasses from @Image 1.
[Stage 1]
Initial state: the model wears the sunglasses against a minimal studio backdrop.
Primary event: a slow camera orbit around the model’s face, light sweeping across the lenses.
End state: profile close-up, the sunglasses sharp in frame.
[Stage 2]
Continue from the previous stage: same model, same sunglasses.
Primary event: slow-motion macro passes over the frame while small elegant white texts appear, connected by thin white lines pointing to the parts they describe: 【High quality steel】 pointing to the hinges, 【Flexible and Resistant】 pointing to the temples, 【Modern design】 pointing to the frame front, 【Sun protection】 pointing to the lenses, 【Iron details】 pointing to the metal accents.
End state: the last text fades while the camera holds on the lenses.
[Stage 3]
Primary event: a distorted transition into a final beauty shot of the model lowering the sunglasses slightly.
End state: the model looks into the camera over the top of the frame.
[Maintain Consistency]
Keep the sunglasses identical to @Image 1 in every shot: same frame shape, materials and colours. Render each on-screen text exactly as written.
Elegant studio light, shallow depth of field, distorted transitions between shots.
(Deep, soft ambient electronic music)
Summer film
VIDEO PROMPT
[Generation Goal]
Generate a short analog summer film set in Andalucía, southern Spain: a warm montage of everyday summer scenes.
[Stage 1]
Primary event: kids play football on the beach in the late afternoon.
End state: the ball rolls toward the shoreline.
[Stage 2]
Primary event: a whitewashed village church stands against a deep blue sky, swifts crossing the frame.
End state: a static hold on the bell tower.
[Stage 3]
Primary event: old women sit on white plastic chairs outside their homes in a whitewashed Andalucían street, chatting and fanning themselves.
End state: one of them laughs, her fan half open.
[Stage 4]
Primary event: a closing montage of small summer details: melting ice cream, bare feet on hot tiles, laundry moving in the breeze.
End state: the sun flares into the lens.
[Maintain Consistency]
Keep the same analog film look and the same warm summer light across all stages.
Analog 8mm film aesthetic: visible grain, warm faded colours, light leaks, slightly unstable handheld framing, analog cut transitions between scenes.
<Film projector rattle under the cuts> <Waves, kids’ voices and cicadas in the distance>
Shaved horse
VIDEO PROMPT
[Generation Goal]
Generate a video where a cowboy shaves a hidden design into a horse’s coat, revealed only at the end: the text “Seedance 2.5” surrounded by stars.
[Stage 1]
Initial state: a horse stands alone in the middle of an open field.
Primary event: a cowboy walks toward the horse holding a razor.
End state: the cowboy stands beside the horse’s flank, razor in his right hand.
[Stage 2]
Continue from the previous stage: same horse, same cowboy, same field.
Primary event: the cowboy shaves patterns into the horse’s coat. The camera stays on angles that keep the design hidden.
End state: the cowboy lowers the razor and walks away; the horse stands alone, the design still unseen.
[Stage 3]
Primary event: the camera moves around to reveal the horse’s flank. Shaved into the fur is the text “Seedance 2.5” surrounded by stars.
End state: the horse rears on its hind legs. <The horse neighs>
[Maintain Consistency]
Keep the same horse, the same cowboy and his clothing, and the same field in every stage. Render the shaved text exactly as written: “Seedance 2.5” with stars around it.
Different camera angles across the stages, from wide field shots to close-ups on the razor. Natural daylight, western film grade.
<The razor buzzes> <Wind moves over the grass>
How to write Seedance 2.5 prompts
Every prompt above is built from the same four moves, straight out of BytePlus’s official spec. Learn them once and you can direct anything.
The core formula
Subject and action are the only mandatory parts. Everything else is a menu, and every part you drop is a decision the model makes without you:
PROMPT FORMULA
<Subject> performs <primary action or event> in <scene and environment>.
The visuals feature <visual style>.
Use <shot size, camera angle, camera movement, or cuts>.
Audio includes <dialogue, ambience, sound effects, or music>.
Style is not a mood word. “Cinematic” leaves the model free. “Analog 8mm film aesthetic: visible grain, warm faded colours, light leaks” does not. That line is what gives the summer film its entire look.
Bind every reference by name
Uploading references? Say exactly what each one contributes, and what it must not contribute. The sneaker and sunglasses prompts open this way:
REFERENCE BINDING
@Image 1 defines the sneakers’ structure, materials and colours.
Do not use the image background. Do not use any person in the image.
The most common failure is writing “@Images 1 through 4 define four characters respectively”. That never says which image is which character. Map them one by one, and write the exclusions every time: a reference with a background you don’t want will hand you that background anyway.
Audio has its own syntax
Four brackets, four channels. You’ve seen all of them in the prompts above:
- ( ) for music. Mood, tempo, instrumentation: (Deep, soft ambient electronic music)
- < > for sound effects, one per bracket: <The razor buzzes>
- { } for dialogue, with language and delivery named before the line: Dialogue language: American English. A deep voice-over says slowly: {The last battle… is now}
- 【 】 for burned-in on-screen text: 【The Final Race】
For 30-second videos, think in stages
Long generations hold best when the story is divided into consecutive stages, each with one primary event and a visible end state. Here is the difference it makes. This was our first draft of the shaved-horse prompt:
FIRST DRAFT
A 30 second video, with different camera angles. A horse in the middle of a field, a cowboy walks towards the horse holding a razor. The man starts shaving the horse, making drawings on its hair with the shaving patterns. We don’t see what the drawings are until the cowboy walks away. We see the horse standing; on its fur is the text “Seedance 2.5” with stars around it, made by shaving its hair.
It worked, and that says a lot about the model’s instincts. But the reveal, the star placement and the final shot were all its decisions. The staged version in the prompt list above takes those decisions back: one event per stage, an end state the camera must land on, and a consistency block that locks the horse, the cowboy and the exact shaved text. Loose prompts are for exploring. Staged prompts are for directing.
Want more ready-to-run prompts? We also built the ultimate Seedance 2.5 animation prompt guide: ten animation styles, every prompt included.
Seedance 2.5 FAQ
What is Seedance 2.5?
Seedance 2.5 is ByteDance’s latest AI video generation model. It generates up to 30 seconds of continuous video from a single prompt, supports text-to-video and reference-based generation, and renders on-screen text with high accuracy. It is available on Magnific.
How long can a Seedance 2.5 video be?
Up to 30 seconds in a single generation, with no stitching or extensions needed. For long videos, structuring the prompt in stages, each with one primary event and a visible end state, keeps the story consistent. Duration is set on the generation page, not in the prompt.
Can Seedance 2.5 generate readable text inside videos?
Yes, and it is one of the model’s strongest capabilities. Specify the exact wording in the prompt and, for burned-in graphics, write it inside 【 】 brackets. Text left to the model’s imagination is where errors appear.
How many reference images can I use with Seedance 2.5?
Up to 30 images and 50 reference materials in total. One reference already goes far for product consistency; more references improve it further. Bind each reference explicitly in the prompt and state what not to use from it.
Where can I use Seedance 2.5?
Seedance 2.5 is available now on Magnific’s AI video generator.
Run your own tests
Six tests told us where Seedance 2.5 bends and where it doesn’t. The seventh test is yours. Take a prompt from this post, change one variable, and see what moves.
Seedance 2.5 is available now on Magnific. START MAKING.
