Seedance 2.5 shipped on July 31, 2026, and ByteDance published the prompts behind its demo clips. They are below, translated — and they are worth studying, because 2.5 changed how prompts are written: shots addressed by timestamp, references addressed by number (@image 1, @video 1, @white-model 1), and up to 50 assets in one request. Our own eight one-take stress tests follow.
Concert, one continuous takeT2V30-second narrative in a single generation
One continuous handheld gimbal follow. The camera pushes slowly through a gap in heavy red curtains into a warm-toned backstage dressing room. A young female singer, back to camera, adjusts her in-ears; a crew member tells her she is on. She turns to look at the camera and begins to sing City Pop. The camera retreats, following her through the curtain into the backstage corridor, where she interacts naturally with her dancers and a crew member hands her a microphone. She and the dancers step onto the stage; the camera orbits behind them as the red-and-black set, LED wall, follow-spots, haze and reflective floor open up. The camera finally pulls out to a full arena view — a packed house, light boards, glow sticks and cheering — a young, free concert-climax atmosphere.
What to copy: The whole arc — dressing room, corridor, stage, arena — is one prompt and one generation. Note that it is written as a journey with handoffs, not as a description of a scene.
Extending a finished clipR2VMulti-round extension into minutes
Extend the video. Continue from the content and subjects of @video 1 and generate another 30 seconds, keeping the characters, setting, visual style and sound effects consistent. A small boy runs along the carriage clutching a football; when the train stops the side doors open and he bolts out, the man chasing close behind. The two run across the platform and out into the street, startling passers-by, until the man finally catches him. The boy looks up, aggrieved; the man's anger fades, he ruffles the boy's hair and gives a resigned smile.
What to copy: The extension prompt names what must stay constant (subject, setting, style, sound) before it says what happens next — that ordering is the whole trick.
Peking opera, orbiting the water sleevesR2VTimestamped shot design with per-character references
16:9 widescreen, cinematic texture, one continuous take, smooth camera movement, no cuts. Scene reference @image 4. 0–5s: open on a close shot of the Xiang Yu character from @image 2, the camera orbiting slowly across his upper body and easing out to a mid shot; he spins and turns, then his body and back-banners sweep past the lens, forming a natural wipe as the camera carries round to Consort Yu from @image 1. 6–10s: a steady orbit around Consort Yu in mid shot, following the water-sleeve movement — she raises an arm, flicks the wrist, opens the sleeve, half-turns, then gathers the sleeve and holds, looking back towards Xiang Yu. 11–20s: the martial role from @image 3 enters with a backflip; Xiang Yu takes centre while the martial role advances and retreats opposite him, Consort Yu behind and to his side answering with her sleeves — hard against soft. The camera pulls slowly from a mid-close on the martial role out to a wide of the stage, ending with all three facing the audience in a synchronised closing pose.
What to copy: The clearest example of the new syntax: a timeline in seconds, and each beat naming which reference image supplies which performer.
Orchestra, eighteen references in one frameR2VUp to 30 images / 10 videos / 10 audio per request
A 30-second concert sequence, 16:9, cinematic realism, real concert-hall light, warm gold stage wash, formal classical atmosphere. Venue @image 1; pianist @image 2; cello @image 3; violin @image 4; lead vocalist @image 5; the rest of the orchestra @images 6–10; the choir @images 11–14; the audience @images 15–18. The vocalist walks downstage centre, the pianist is at the piano, the orchestra spreads to either side and behind, the choir stands upstage. Open on a high overhead wide of the hall; the pianist strikes the first chord, the vocalist steps into the follow-spot and sings, the camera drifting naturally across violins, cellos and the orchestra — the violins bright, the cellos warm. Later the choir joins; the vocalist and someone in the front row exchange a brief look and the audience member smiles and nods. The camera pulls back as the song ends and the audience applauds.
What to copy: Eighteen separate references, each holding its own identity in the same frame. This is the prompt that justifies the jump from 9 assets to 50.
Greybox layout, renderedR2VWhite-model reference for blocking and lighting
Take the camera movement, cutting rhythm, shot sizes, subject trajectories and camera blocking from @white-model 1. Take the character design, setting, materials, lighting, colour and fairy-tale mood from @image 2. Render the greybox as a dreamlike, warm, children's-fantasy 3D animated short. Beats in order: flying through a fantasy sky → sea-of-cloud spirit beasts flying alongside → diving into the ocean → a manta ray weaving through the deep → a mirrored rift in time → picking a star in space → transforming back into the bedroom → father tucking in the covers → the picture book closing on a final frame.
What to copy: Untextured 3D as the blocking layer: the greybox carries camera and motion, the reference image carries look. Lighting is derived from the greybox's spatial information rather than invented.
Green screen, three worldsR2VGreen-screen editing with per-scene physics
In @video 1, render the green-screen background into scenes, obstacles, costumes and supporting characters. 0–4s: outdoor training — replace the obstacles with rocks, bricks, tyres and wooden crates. 4–10s: a locker room, friends offering encouragement. 10–15s: an international stadium — replace the training poles with original defenders and a goalkeeper; the protagonist scores. Realistic, cinematic throughout.
What to copy: The subject is preserved while the world changes three times, and the model is expected to re-derive cloth movement, hair, gait and light for each new environment.
Re-shooting the camera, keeping the performanceR2VCamera-move editing by timestamp
Edit @video 1. Keep the characters, action and visual style unchanged; adjust only the camera. A 15-second camera plan in segments: 0–4s, a micro FPV skimming the pan, following the toast as it pops up then whipping sideways to the coffee; 4–7s, push in and track laterally along the pan edge, following the egg as it flips and lands; 7–11s, rise rapidly to a top-down view and press down evenly across the plates and keys; 11–15s, handheld close following two hands in a fast lateral sweep, finally pushing into the breakfast and pulling back to a two-shot. Continuous and stable throughout.
What to copy: Editing addressed purely at the camera: the performance is fixed, the coverage is rewritten. Closer to a director's note than a prompt.
Classroom: a Song-dynasty streetR2VEducation use — Doubao Classroom is already shipping this
Eastern freehand-ink style. A street scene in Southern Song Lin'an; on the bustling street several children run and play, chanting the line “I turned, and there she was, where the lanterns were growing dim.” The camera stays with the running children, sweeping past the busy street, then tilts up: this man is Xin Qiji, from @image 1. Xin Qiji turns his head; in the distance a figure stands in the fading lantern light. Continuous camera throughout.
What to copy: A literature lesson as a shot: the classical line is spoken in-scene, and a single reference image supplies the historical figure.
Car assembly from greyboxR2VIndustrial simulation and synthetic data
Take the camera movement, composition, shot sizes, spatial relationships, part positions, model structure, assembly order and motion trajectories from @white-model 1. Take the materials, lighting, colour, reflections and atmosphere from @image 1. Render the greybox as a high-end, photorealistic car-assembly sequence.
What to copy: The same greybox technique pointed at manufacturing rather than storytelling — this is the synthetic-data pitch in one prompt.
Official prompts published by ByteDance's Seed team with the Seedance 2.5 launch, July 31, 2026. Translated from Chinese; the reference and timestamp syntax is kept verbatim. The demo clips are hosted inside WeChat and can't be mirrored here.
Our eight one-take stress tests
Written before launch for Seedance 2.5's headline capability — one continuous 30-second take. They push it deliberately: water transitions, reverse physics, style shifts, lighting changes mid-shot. The API is not open yet, so they are still untested; the day it opens we render all eight and publish the outputs here, unedited.
Ramen kitchen pass-throughContinuous handoff
One continuous 30-second take, no cuts. The camera drifts through rising steam above a boiling ramen pot, tilts down as the chef's ladle pours golden broth into a bowl, then tracks the bowl as it slides along the pass-through counter — chashu, egg and scallions landing in rhythm as it moves. The camera keeps following over the counter and settles on a customer by the window taking the first slurp, chopsticks lifting noodles high, while neon-lit rain streaks the glass behind them. Warm tungsten kitchen light giving way to cool neon window light. Audio: rolling boil, ladle pour, bowl sliding on wood, chatter, rain against glass.
Why this tests 2.5: Object handoff and steam physics sustained across three spaces without a cut — segment-stitched models drift the bowl's contents.
A single slow lateral dolly across a studio apartment, one unbroken 30-second shot. As the camera moves, the world outside the window cycles through the seasons: spring rain, summer glare, red autumn leaves, then snow. Inside, the room ages in sync — a potted plant grows taller and drops a leaf, books accumulate on the desk, the blanket on the chair changes from linen to wool, and the light temperature shifts from cool to golden to blue-gray. No people. The dolly never stops, never cuts. Audio: a soft piano loop that changes key with each season, muffled weather sounds through the glass.
Why this tests 2.5: Thirty seconds of continuous scene-state mutation — everything must transform coherently while the camera never looks away.
One unbroken 30-second Steadicam shot. Start close on a singer's face in a dim dressing-room mirror ringed with warm bulbs; she exhales, turns, and walks — the camera swings behind her shoulder and follows down a narrow concrete corridor, past crew members flattening against the wall, cables snaking underfoot, the muffled crowd noise growing with every step. A stagehand pulls open a heavy door: blinding white stage light floods the frame, the roar snaps to full volume, and the camera rises past her silhouette to reveal a stadium as she lifts the microphone and sings the first note. Audio: heels on concrete, breathing, crowd swell from muffled to deafening, one clear sung note to end.
Why this tests 2.5: A hard lighting-regime and acoustic transition mid-shot — the classic 'oner' every cinematographer knows, and a brutal continuity test.
One continuous 30-second tracking shot chasing a scooter courier with a red insulated box through a packed Taipei night market. The camera weaves with him between food stalls — through billowing grill smoke, under a row of hanging lanterns that brush the lens, past a vendor tossing noodles in a wok flare — ducks beneath a fabric awning, splashes through a neon-reflecting puddle, and arrives as he brakes, dismounts in one motion, and hands the box to a grandmother in a doorway who is already reaching out. Handheld energy, never losing him behind the crowd. Audio: scooter engine weaving, wok sizzle passing left to right, market chatter, brake squeak, a soft 'xièxie'.
Why this tests 2.5: Dense crowds, smoke, and repeated occlusions with one persistent subject — where identity drift usually happens.
Reference the sneaker from Image 1. One continuous 30-second orbit in a white infinite studio void: components rain down in slow motion — sole, upper, laces, eyelets — and assemble themselves in mid-air exactly into the sneaker from Image 1, each part clicking into place as the camera circles. The orbit tightens as assembly completes; the finished shoe rotates once, drops onto a lit plinth with a soft bounce, and the camera comes to rest in a low hero angle. Studio softbox lighting, sharp shadows under the plinth. Audio: whooshes on each falling part, tactile clicks and lace-zips, one deep bass hit on the plinth landing.
Why this tests 2.5: Reference fidelity held through 30 seconds of parts-in-motion — the product must land identical to Image 1 from every angle.
A fixed slow orbit around a marble table, one 30-second take, with time flowing backward for everything except the camera. A demolished chocolate cake reassembles: crumbs leap from the tablecloth back into clean edges, a pool of raspberry sauce retreats up into a hovering spoon, a bitten strawberry un-bites to whole, candle wax un-melts as its flame shrinks to a spark, and finally a chef's hands enter in reverse to un-place the plate and back away. The orbit never changes speed. Elegant restaurant dusk light. Audio: a reversed piano piece resolving into silence, subtle reversed foley — clinks and pours played backward.
Why this tests 2.5: Reverse-time physics under a forward-moving camera — coherence here is a genuine generation milestone, and instantly shareable.
One unbroken 30-second shot in a quiet museum at closing time. The camera pushes slowly toward a large oil painting of a 19th-century harbor at sunset; as it crosses the frame the world turns real-but-painterly — visible brushstrokes on the waves, impasto clouds, a fishing boat rocking as gulls wheel overhead. The camera sails between two ships, turns, and pulls back out through the frame into the gallery, revealing a security guard asleep in his chair and one alarm light beginning to blink red on the wall. Audio: silent gallery hum, then wind, rope creak and gulls inside the painting, then the hum again with a soft periodic alarm chirp.
Why this tests 2.5: Two full style transitions — photoreal, painterly, photoreal — inside one camera move, with the gallery state advancing while we were 'inside'.
One continuous 30-second shot at first light. A drone-smooth camera chases a swimmer sprinting barefoot down a weathered wooden dock; she dives, and the camera plunges into the water with her — the frame submerging in one seamless motion into blue-green silence. It follows her glide through a kelp forest, silver fish scattering, sunbeams cutting down in shafts, then rises with her as she surfaces beside a red buoy exactly as the sun breaks the horizon and floods the frame gold. Audio: footfalls on wood and gulls, the plunge muffling everything to underwater ambience and heartbeat, then a gasp of air, lapping water and full birdsong.
Why this tests 2.5: The above-to-below-to-above water transition in one take — a shot real filmmakers need housing rigs for, and a known failure point for video models.
One email when Seedance 2.5 ships, with all eight prompts rendered side by side — what held together and what fell apart. No other emails.
Until then: what works on Seedance today
Seedance 2.0's real strength is multi-modal referencing — prompts that direct with "Image 1" and "Video 1" inputs. Our Seedance prompt library has official prompt-output pairs you can copy now, and the Seedance prompt generator builds prompts in the three-block prose format the model prefers. Planning a multi-shot piece instead of a one-take? Use the Shot Planner.
Frequently asked questions
Has Seedance 2.5 been released?
Not yet. ByteDance previewed it on stage on June 23, 2026; the expected window is late July to early August 2026. The only officially confirmed capability is 30-second single-shot generation — which is exactly what these prompts are written for. Our Seedance 2.5 model page tracks what's confirmed versus rumored.
Have these prompts been tested?
They can't be tested on 2.5 until it ships — nobody outside ByteDance has access, and we won't pretend otherwise. Each prompt is designed as a stress test of what a true 30-second one-take must hold together (continuity, state persistence, lighting transitions). On launch day we render all eight and publish every output here, unedited, successes and failures both.
Can I use them on Seedance 2.0 today?
Yes. Seedance 2.0 builds long videos by extending segments, so split any of these prompts at its natural seams — the door opening, the dive, the frame crossing — and generate each part with the previous clip as reference. The one-take magic is what 2.5 promises to remove that workaround.
Why one-take prompts specifically?
Because 30-second single-shot generation is the only thing ByteDance has confirmed, and it's the structural difference from 2.0. A continuous take can't hide errors behind cuts: identity drift, physics resets and lighting jumps are all exposed. If 2.5 delivers, these prompts show it off; if it doesn't, they'll show that too.