The prompts you write decide whether a clip is ready to publish or needs five regenerations. This Seedance 2.5 prompt guide breaks down the structure, camera syntax, and examples you can paste straight into the model.
Quick answer — the formula
Subject + Action + Camera + Lighting + Style (+ Audio)
A prompt that names all five layers gives Seedance 2.5 the scaffolding it needs to keep a long shot coherent. Drop any layer and the model fills the gap with a guess.
The formula in one sentence
A young barista in a green apron steaming milk behind a marble counter, turning to smile at the camera. Slow dolly-in from a wide shot to a medium close-up. Warm morning light through a side window, soft shadows. Cinematic, shallow depth of field, 35mm film look. Ambient café noise and the hiss of the steam wand.
Every strong Seedance 2.5 prompt is built from the same five ingredients. Think of them as layers you stack rather than a rigid order to obey.
Vague subjects produce generic, unstable footage. Don't write "a woman"; write "a woman in her thirties with short dark hair, wearing a beige trench coat." Concrete attributes — age, clothing, color, material, expression — give the model anchors to hold steady across a long clip. For products, name the object, its material, and its finish.
A long clip can hold a complete beat, so describe an arc with a beginning and end: "she picks up the cup, takes a sip, and sets it down." Avoid stacking unrelated actions; the model handles a single continuous motion far better than a list of jump-cuts.
Camera language is where most prompts are too thin. Specify the shot size (wide, medium, close-up), the angle (eye-level, low angle, overhead), and the movement (static, dolly-in, pan, orbit). Seedance 2.5 follows camera direction well, so use it to add production value.
Lighting carries the emotional tone. "Golden hour backlight," "soft overcast daylight," "moody neon from the left," and "high-key studio lighting" each produce a completely different feel from the same subject. Because the output carries rich color, lighting detail actually survives into the final file.
Close with the aesthetic: "cinematic, shallow depth of field, 35mm film grain," "clean commercial product look," "documentary handheld," or "anime cel-shaded." Style words tie the whole frame together and keep the model from defaulting to a flat, in-between look.
Camera moves are the fastest way to make AI video look intentional. Use plain film terms — Seedance 2.5 understands them — and pair each with a speed word ("slow," "smooth," "fast"). Combine at most one or two moves per clip.
Camera moves toward the subject. Great for reveals and emphasis: "slow dolly-in to a close-up."
Camera retreats to reveal context: "pull back to reveal the full room."
Camera pivots horizontally: "smooth pan right across the skyline."
Camera pivots vertically: "tilt up from the shoes to the face."
Camera circles the subject: "slow 180-degree orbit around the car."
Camera rises or descends: "crane up from street level to a rooftop view."
Camera moves alongside a moving subject: "tracking shot following the runner from the side."
Adds subtle, natural shake: "handheld documentary feel, slight camera movement."
No movement, for clean product or interview shots.
Written to take advantage of the long single-shot duration, high resolution, and native audio. Paste them in, then swap the details for your own subject.
A matte black wireless speaker rotating slowly on a white pedestal, water droplets beading on its fabric grille. Slow 360-degree orbit, then a gentle dolly-in to the logo. Clean high-key studio lighting, soft reflections. Premium commercial product look, crisp detail. Subtle ambient electronic hum.
A barista in a denim apron pouring latte art into a ceramic cup, steam rising. Static medium shot, then a slow push-in to the cup. Warm morning window light, soft shadows. Cozy lifestyle aesthetic, shallow depth of field. Ambient café chatter and the clink of a spoon.
A detective in a long coat steps out of the rain into a dim doorway, water dripping from the brim of his hat, looking off-screen. Slow dolly-in from a wide shot to a medium close-up. Moody blue night light with a single warm streetlamp. Noir cinematic look, film grain. Rain ambience and distant thunder.
A chef's hand drizzling olive oil over a fresh burrata salad on a slate plate, basil leaves glistening. Overhead static shot tilting to a low angle. Soft natural daylight from the left. Appetizing food-commercial style, rich color, shallow focus. Gentle kitchen ambience.
A bright modern living room with floor-to-ceiling windows and minimalist furniture. Smooth forward tracking shot gliding from the entryway through to the balcony view. Natural midday light, airy and clean. Architectural-tour look, sharp detail. Quiet ambient room tone.
A smartphone floating against a soft gradient background, its screen animating through a clean fitness app interface. Slow dolly-in toward the screen, gentle parallax. Bright soft studio lighting, subtle reflections. Modern tech-commercial style, vivid color. Light UI tap sounds and a soft synth pad.
Morning mist drifting over a pine forest valley as the sun breaks through the trees. Slow crane up revealing the layered mountain ridges beyond. Warm golden-hour backlight, volumetric god rays. Cinematic documentary look, deep detail. Birdsong and a soft breeze.
A model in a flowing red dress turning under studio lights, fabric catching the air. Smooth half-orbit around the model, then a slow tilt up. Dramatic high-contrast lighting with a dark background. Editorial fashion-film look, shallow depth of field. Subtle rhythmic music bed.
A friendly presenter in a casual blazer speaking directly to camera in a bright home office, gesturing naturally. Static medium shot with a very slow push-in. Soft key light from the front, warm and even. Clean explainer look. Clear spoken dialogue: "Here's the one setting that changes everything."
A sleek electric SUV driving along a coastal cliff road at sunset, dust trailing behind the wheels. Tracking shot following the car from the side, then a slow pull back to reveal the ocean. Warm golden light, long shadows. Premium automotive-ad look, cinematic. Engine hum and wind, building music.
When you upload a still, the image already defines the subject, composition, and look — so your prompt's job is to describe motion and camera, not to re-describe what's already visible.
Focus on what should happen
The subject turns their head toward the camera and smiles. Gentle breeze moves the hair. Slow dolly-in. Keep the existing lighting and composition.
Don't repeat "a woman in a red dress" if she's already in the photo — say what she does.
Add "keep the existing lighting and composition" when you want the model to honor the still rather than reinterpret it.
A calm portrait calls for subtle motion; an action frame fits a bigger move. Mismatched motion is the most common reason a result feels uncanny.
A single believable movement beats a complicated choreography that fights the source image.
Seedance 2.5 accepts multiple multimodal references in a single generation — images, audio clips, 3D white models, and style references. This is the model's superpower for anyone producing more than one clip, and your prompt should work with the references, not against them.
Attach reference images of your character, then write "the character from the reference images" instead of re-describing them. This keeps identity stable across multiple shots.
Attach a style reference and prompt "in the visual style of the reference," then describe your new subject and action normally.
In motion-reference mode, attach a clip whose movement you want to reproduce and describe your new subject performing that motion.
Reference your product shots and prompt for "brand-consistent styling and colors from the references."
References lock identity and look; the prompt directs action and camera. When you catch yourself re-describing something already in a reference, delete it and point to the reference instead.
Because Seedance 2.5 generates sound in the same pass as the picture, audio belongs in your prompt — not as an afterthought. Add a short audio clause at the end.
"ambient café noise," "ocean waves and wind," "quiet room tone."
"footsteps on gravel," "the hiss of a steam wand," "a car engine starting."
Put the exact line in quotes — spoken dialogue: "Let's get started." Keep lines short for the best lip-sync.
"soft synth pad," "building cinematic score," "upbeat pop bed."
One clear ambience plus one optional effect or line is plenty. Over-stuffing the audio clause makes the mix muddy. If you don't mention audio at all, the model still adds fitting sound — but you lose control over what it picks.
"A nice video of a city" gives the model nothing to hold onto. Name the subject, time of day, and camera.
Three unrelated motions in one clip produces stutter and morphing. Describe one continuous arc.
"Static locked-off shot with a fast orbit" cancels itself out. Pick one camera behavior.
Repeating what's already in your uploaded still confuses the model. Describe only the motion.
These two add the most production value for the least effort — skipping them is why a prompt looks flat.
With native sound available, leaving audio to chance wastes a built-in advantage.
"Beautiful stunning gorgeous amazing cinematic masterpiece" doesn't help. Concrete nouns and verbs beat stacked adjectives every time.
Before you publish a generation, run this quick check — and regenerate with a tweaked prompt if anything fails. A failed check is usually a prompt problem, not a model problem.
Does the character or product stay identical from the first frame to the end? If it drifts, add references or sharpen the subject description.
Is the movement smooth and physically believable, with no morphing or warping? Simplify to one action if not.
Did the model execute the shot you asked for? If it ignored the move, state it more plainly and remove competing directions.
Do sound effects and dialogue line up with the on-screen action? Shorten dialogue lines for better lip-sync.
Is the output clean in close-ups, with no smeared textures?
For commercial work, check text, logos, and anything on-screen that needs to be accurate before it goes live.
Name the subject, give it one clear action, choose a camera move, set the lighting, lock the style, and add the audio you want. Then generate your first cinematic clip.
