Cheat Sheet · Prompting & Keywords
Seedance 2.0 takes up to 12 reference files per generation, 9 images, 3 video clips and 3 audio tracks, yet most people type one plain sentence and use maybe 15% of the tool. This guide is how to use Seedance 2.0 properly, the reference we keep open while generating: the 5-layer prompt structure, the camera and lighting keywords that actually land, and the constraints that keep output stable.
The official documentation lists a 6-element formula, but community testing settled on 5 layers that consistently beat longer, looser prompts. The order carries weight, each layer locks something down before the next one adds to it.
1 · Subject | Who the shot is about. Specific identity markers (hair, clothing, accessories) pin the model, anything you skip gets averaged into something generic. |
2 · Action | One primary movement, present tense. Write directions, not states, "she slowly turns toward the camera" beats "she looks happy". |
3 · Camera | One primary camera movement, described by rhythm (slow, smooth, gentle) rather than f-stops, ISO values or millimeter specs. |
4 · Style | Lighting, color grade and film references. Late in the stack on purpose, it adds visual flavor without hijacking the motion. |
5 · Constraints | Guardrails like "avoid jitter" and "maintain face consistency" that close whatever gaps the first four layers left open. |
Example prompt
a man in his 40s, weathered features, worn denim apron ← subject
he carves a walnut board, wood shavings curling off the blade ← action
slow push-in from medium shot to close-up on his hands ← camera
cinematic film tone, 35mm, warm golden hour light ← style
avoid jitter, avoid bent limbs, stable picture, no flickering ← constraintsThis is where most prompts collapse, people write feelings instead of directions. Two habits fix it: one clear mover per clause, and describing what happens instead of how it should feel.
One subject per shot | The safest path. Two characters work if you separate them spatially and tag them @Character_A and @Character_B, three or more rarely holds together. |
Stack identity markers | "a woman in her late 20s, tight dark curls at ear length, small silver hoop in left ear, fitted black turtleneck", every detail you specify is one the model doesn't hallucinate. |
Directions, not states | "she slowly turns toward the camera, breeze lifting the hem of her skirt" gives the model a sequence to execute; "she is enjoying the sunset" gives it a photograph to approximate. |
Split subject & camera motion | "the dancer spins slowly, camera holds fixed framing", one instruction per mover eliminates most of the shaky output people blame on the model. |
Example prompt
✗ a woman enjoying the sunset on a beach
✓ a woman in her late 20s, tight dark curls at ear length, small
silver hoop in left ear, fitted black turtleneck, she slowly
turns toward the camera, breeze lifting the hem of her skirt,
camera holds fixed framingSeedance treats camera controls as a first-class signal, which is where it separates from other video models. Pick one primary movement per generation and describe its rhythm, not its hardware.
fixed / locked-off | Zero camera movement; add "locked tripod, zero camera shake" if ambient jitter persists |
static wide | Wide, unmoving establishing shot |
push-in / dolly in | Camera moves toward the subject, tension, emphasis, emotional close-ups |
pull-out / dolly out | Camera moves away, environmental reveals and context |
pan left / pan right | Horizontal rotation in place, scanning or following action |
tracking shot / follow | Camera moves alongside the subject, action sequences |
orbit / arc / 360 orbit | Circles the subject, product showcases, portraits, hero moments |
aerial / drone shot | High altitude, landscapes and establishing geography |
handheld | Natural shake, documentary feel, UGC authenticity |
crane up / crane down | Vertical ascent or descent, dramatic height reveals |
gimbal | Smooth stabilized motion, polished cinematic, distinct from handheld |
steadicam walk | Smooth forward motion following a character through space |
whip pan | Rapid horizontal sweep, urgency and scene transitions |
dolly zoom | The Hitchcock vertigo effect, subject stays the same size while the background warps |
rack focus | Shifts focus between foreground and background planes to redirect attention |
Example prompt
a barista pours latte art in a sunlit café, steam rising from
the cup, orbit shot circling the counter slowly, shallow depth
of field, background softly blurred"fast" is the most dangerous word in a Seedance prompt, fast camera plus fast subject plus a busy scene almost guarantees jitter and compression artifacts. Make one element fast and hold everything else steady.
imperceptible / barely | Extremely slow, almost unnoticeable movement |
slow / gentle / gradual | The default recommendation and the safest starting point |
smooth / controlled | Natural rhythm |
dynamic / swift | High impact, use with caution and only on a single element |
start: …, then: … | Sequence compound moves instead of stacking them: "start: slow dolly-in, then: gentle pan right for the final 2 seconds" gives two temporal phases, not two competing instructions |
Example prompt
a boxer wraps her hands in a quiet gym, only her hands move fast,
start: slow dolly-in from wide to medium,
then: gentle pan right for the final 2 seconds as she looks up,
smooth controlled motion throughoutPer the official guide, lighting has the single biggest impact on quality of any prompt element, more than style adjectives, quality modifiers or resolution requests. If a weak prompt gets one addition, make it a lighting line.
golden hour | The highest quality-per-word upgrade you can add |
rim light against dark background | Cinematic edge separation |
soft key from 45 degrees | Flattering talking-head lighting |
overcast daylight / even diffused light | Eliminates flicker in bright scenes |
backlit silhouette at sunset | Dramatic mood |
motivated lighting from practical source | Realism with the light source visible in frame |
volumetric fog | Atmospheric depth, pairs well with backlit setups |
chiaroscuro | High-contrast lighting in the Godfather mold |
Example prompt
an elderly fisherman mends a net on a wooden pier, golden hour,
dramatic rim light against the dark water, volumetric fog rolling
in over the sea, motivated lighting from a lantern beside him"Cinematic" on its own is too vague to produce anything predictable. Pair it with a texture, a grade or a film stock so the model gets intersecting constraints instead of permission to improvise.
cinematic film tone, 35mm | The most reliable all-purpose anchor |
16mm film, handheld camera | Raw indie aesthetic |
anamorphic lens flare | Widescreen cinematic |
national geographic quality | Nature documentary |
documentary-style handheld framing | Observational realism |
teal and orange | Classic Hollywood grade |
bleach bypass | Desaturated, gritty, high-contrast texture |
warm tone / amber-tinted | Nostalgic feel |
crushed blacks | Deep cinematic shadow loss |
pastel | Soft anime or fashion aesthetic |
Example prompt
two friends ride bicycles through a rain-washed street at dusk,
cinematic film tone, 35mm, teal and orange grade, crushed blacks,
anamorphic lens flare from passing headlightsConstraints are the layer that separates AI-looking output from video that passes. Positive statements ("avoid X", "maintain Y") read more reliably than negative-prompt syntax, so phrase every guardrail that way.
avoid jitter | Prevents screen shaking |
avoid bent limbs | Prevents distorted arms and legs, use in every character prompt, no exceptions |
avoid identity drift | Prevents character features changing between shots |
avoid temporal flicker | Prevents frame-to-frame brightness oscillation |
no distortion, no stretching | Maintains geometric stability |
maintain face consistency | Preserves face identity across cuts |
sharp clarity, natural colors, stable picture, no blur, no ghosting, no flickering | The community quality suffix, inelegant and measurably effective, append it to every generation |
Example prompt
a dancer leaps across an empty stage under a single spotlight,
slow motion, static wide shot,
avoid jitter, avoid bent limbs, maintain face consistency,
sharp clarity, natural colors, stable picture, no blur,
no ghosting, no flickeringThe pattern under all of these: if a word describes how the viewer should feel instead of what the camera should see, the model has to guess which visual produces that feeling, and it guesses wrong.
"fast" (unqualified) | Accelerates everything at once. Name which single element moves fast and hold the rest steady |
"cinematic" (alone) | Gives the model nothing to work with, always pair it with texture, lighting or a film reference |
"epic" | No visual meaning to a diffusion model |
"amazing" / "beautiful" / "stunning" | Feelings, not instructions, the model can't render an adjective |
"lots of movement" | Triggers jitter across the whole frame, name one specific movement instead |
"glow" / "glimmer" / "glints" | Invite specular flicker, write "steady intensity" or "diffuse" instead |
Example prompt
✗ epic cinematic shot, lots of movement, stunning glowing lights,
fast action
✓ tracking shot follows the sprinter, only she moves fast,
stadium floodlights at steady intensity behind her,
cinematic film tone, 35mmA single generation takes up to 9 images, 3 video clips and 3 audio tracks alongside your text, 12 files processed in one pass, this is how image-to-video, motion transfer and character consistency actually work in Seedance. Every upload needs an explicit role in the prompt, an untagged file gets processed ambiguously, and ambiguity averages into mush.
@Image1 … @Image9 | Character sheets, mood boards, product photos, storyboard panels. Tag the role: "character from @Image1 (maintain exact facial features and outfit)" |
@Video1 … @Video3 | Camera motion, choreography or pacing reference: "camera performs slow orbit matching @Video1's motion arc" |
@Audio1 … @Audio3 | Voiceover, music or sound effects: "scene transitions align with beat positions of @Audio1" |
First–last frame | Upload the opening frame as @Image1 and the closing frame as @Image2, describe what happens between them, Seedance interpolates the motion, no storyboarding pipeline needed |
Example prompt
character from @Image1 (maintain exact facial features and outfit)
walks through the environment from @Image2 (match lighting and
color palette), camera performs slow orbit matching @Video1's
motion arc, scene transitions align with beat positions of @Audio1,
avoid identity drift, avoid jitter, stable pictureYou can direct individual shots inside one 4-15 second generation by writing timestamps into the prompt. Give each shot its own camera position, subject action and lighting state, and spell out the transitions.
[0-4s]: … | Range-bracket format: "[0-4s]: wide establishing shot, static camera, misty bamboo forest at dawn, golden hour light" |
(0-3s) … | Parenthetical format, same effect: "(0-3s) macro shot of perfume bottle among pink flowers, shallow depth of field" |
hard cut to / seamless morph into | Explicit transition language between shots, so the model doesn't improvise the cuts |
wide → medium → close-up → extreme close-up | The 15-second climax arc: establish (0-4s), build tension (4-8s), emotional peak (8-12s), reveal or slow-motion hold (12-15s) |
Example prompt
(0-3s) macro shot of perfume bottle among pink flowers, shallow
depth of field, petals floating
(3-7s) camera glides closer, a hand enters frame, touches the bottle
(7-12s) slow-motion spray, mist diffusing, particles catching rim light
(12-15s) pull-out to hero frame, product centered, volumetric
lighting, minimal backgroundThe instinct after a failed generation is to rewrite the whole prompt, subject, camera, style and lighting at once, and then you can never isolate what helped. Controlled iteration is slower per cycle and converges faster, the same reason A/B tests beat redesigns.
2-3 baselines first | Run the same prompt a few times before judging it, diffusion output varies between takes |
One variable per pass | Change the camera, the lighting or the speed modifier, never all three, then score each take for continuity and adherence |
dynamic motion / vibrant energy | Global intensity modifiers, drop one at the start of the prompt when movement is too subtle, they amplify specified motion without adding new movement types |
Keep the best, change one more | Score, keep the winner, adjust a single variable, repeat until the shot holds |
Example prompt
pass 1: golden hour, slow push-in
→ keeper, but the motion reads too subtle
pass 2: golden hour, slow push-in, dynamic motion
→ one variable changed, motion amplified, doneSpin up a hosted open-source AI agent with Creative Tim, bring your own key or run it on credits.
Seedance 2.0 is ByteDance's multimodal AI video model, released in early 2026. It generates 4-15 second clips with synchronized stereo audio in a single pass, and unlike text-and-image-only rivals it accepts four input types at once: text, images, video clips and audio tracks.
Five layers in a fixed order: subject, action, camera, style, constraints. Subject first pins the model to a center of gravity, camera third locks framing before the model re-decides the lens, and constraints last close whatever gaps the other layers left open.
Up to 12 in a single generation: 9 images, 3 video clips and 3 audio tracks, plus your text prompt. Tag every upload with an @ role (@Image1, @Video1, @Audio1), an untagged file gets processed ambiguously and drags the output toward mush.
Usually a speed problem: the word "fast" combined with a moving subject and a busy scene. Let one element move fast and keep the rest steady, add "avoid jitter" and "avoid temporal flicker", and swap "glow" or "glimmer" for "steady intensity" or "diffuse".
Yes. Write timestamps into the prompt, [0-4s]:, [4-9s]: and so on, across up to 15 seconds, and give each shot its own camera, action and lighting. Transition phrases like "hard cut to" control the cuts instead of leaving them to the model.
4 to 15 seconds per generation at up to 1080p, with dual-channel stereo audio generated in the same pass, lip-synced speech across 8+ languages, background music and foley included, not stitched on afterwards.
Generations run from roughly $0.60 each depending on resolution and duration. Seedance 2.0 is available through partner platforms with credit-based plans, several of which include free trial credits, there's no unlimited free tier.
Yes, ByteDance exposes Seedance through the Volcengine / BytePlus platform, so you can call it programmatically with the same prompt structure covered on this page, the 5-layer stack, @ reference tags and time-coded shots all apply.
Compare Seedance 2.0