Guide · Prompt writing

Seedance 3.0 Prompt Guide

Write prompts the way a director writes a shot list. This guide walks through real Seedance 3.0 prompts, the files that went in and the videos that came out — from 30-second one-take shots to video editing, audio translation and fixes for common issues.

one-take video
30s
reference files
50
spoken languages
11
Full HD output
1080p

01

Key capabilities of Seedance 3.0

Seedance 3.0 generates longer takes, accepts far more reference material, sizes the output for you and speaks many languages. Every card below shows the real prompt, the input files and the result.

Text + 8 images → 30s

30-second one-take videos

The maximum length is now 30 seconds (up from 15 seconds in Seedance 2.0), so a whole story fits in a single generation with no clips to stitch together.

Prompt

One continuous take. The camera smoothly follows a person in a black coat (see Image 1) walking from left to right through six connected rooms, each with a different color palette and mood. Every room has the same structure — white walls, light herringbone wood floor, French double windows and white sheer curtains, as in Image 2 — but the view outside and the atmosphere inside are completely different. The person walks at a steady pace through each open doorway. 0–5s: Room one, theme: comic-book fight. The person enters and fights a character (Image 3), who is defeated. 5–10s: Room two, theme: warmth, in a needle-felt style. A sunflower field outside the window (Image 4), warm orange light inside, a painter painting sunflowers (Image 5). The person turns needle-felted on entering. 10–15s: Room three, theme: sadness, in black-and-white comic stop-motion. Rain outside, cold grey light inside; someone sits alone on the floor hugging their knees, a phone beside them showing an unanswered call. The person switches the light off and on again — the room turns to color and fills with flowers. 15–20s: Room four, theme: joy. The room is submerged in the sea (Image 6); the person swims in among coral reefs and schools of fish. 20–25s: Room five, theme: surprise. Fireworks fill the night sky outside (Image 7); colorful flashing light inside; the person is swept up in the cheering. 25–30s: The person reaches an empty white room, stands in the center and snaps their fingers (snap sound effect). The screen cuts to black and the word “seedance” appears in the middle (Image 8). Cinematic, high-end fashion-ad look. Lighting in each room is driven entirely by the view outside, creating strong emotional contrast. No on-screen text.

Inputs
Image 1
Image 2
Image 3
Image 4
Image 5
Image 6
Image 7
Image 8
Output
Text + 1 image + 6 videos

Up to 50 multimodal references

Mix up to 50 files in one request — 30 images, 10 videos and 10 audio clips. You can also drive a video from audio alone, with no image or video needed.

Prompt

Bright, colorful commercial style. The hero is a fruit-flavored biscuit in four flavors — strawberry, apple, grape and orange; the strawberry one follows Image 1. Biscuits and their fruits are arranged in tightly ordered geometric grids; the frame is clean, premium and rhythmic. Open with fruit to grab attention fast, framed like Video 1, with the music hitting on the downbeat. Then the different flavors line up neatly — cut to close-ups, with motion and camera work from Video 2. At the climax a biscuit snaps in half and drops into slow motion: the fruity filling bursts, crumbs fly, the juiciness and texture are magnified, with the impact of Video 3. A horizontal array creates a rhythmic arc, moving like Video 4, highlighting order and variety. Then snap back to fast cuts. At the end, the English line “One bite of crispness, a heart full of delight” enters word by word with punchy text motion and a product freeze-frame, like Video 5. Close on the brand: biscuits and fruit burst outward like Video 6. The whole film feels young, energetic, delicious and made to share.

Inputs
Image 1
Video 1
Video 2
Video 3
Video 4
Video 5
Video 6
Output
Auto ratio · auto length

Smarter duration and aspect ratio

Leave the aspect ratio and duration on automatic and Seedance 3.0 picks sensible values from your prompt and files. Editing, extension and first/last-frame tasks follow special rules:

Video editing

Prompt

Video edit: remove everyone from Video 1 except the main character.

Inputs
Video 1
Output

The output keeps the edited video’s ratio and length: here 5:4 and about 16 seconds. It can come out up to 0.4s shorter than the source.

Video extension

Prompt

Extend Video 1: the window opens onto the gallery interior from Video 2, and finally the camera moves into the painting from Video 3.

Inputs
Video 1
Video 2
Video 3
Output

The output keeps the ratio of the video being extended (7:5 here), even though Videos 2 and 3 are 16:9. You can set the length yourself (4–30s) or leave it on auto.

First & last frame

Prompt

The girl in the picture says “cheese” to the camera; 360-degree orbit shot.

Inputs
Image 1
Image 2
Output

The output matches the first frame’s ratio — 1:1 here.

Text + 1 image → 20s

Native multilingual generation

Write prompts in, and generate speech in, Chinese, English, Spanish, Indonesian, Malay, Thai, Arabic, Portuguese, Vietnamese, Japanese and Korean — with lip-sync that follows each language.

Prompt

Cinematic hip-hop music video, photoreal, premium, on a beach. Build the scene from Image 1: a band plays on golden sand where waves break — a lead vocalist on a mic stand planted in wet sand, a guitarist on the left, another guitarist on the right, a drummer behind a kit. The coastline stretches behind them, waves roll in, a huge warm golden-hour sun skims the sand and glitters on the water, sea mist and salty haze in the air. The lead in a red tracksuit raps to camera with rhythm — mouth and jaw land precisely on every word, head nodding hard on the beat. The band sways in the groove. An upbeat, punchy rap track: fast, confident, heavy beat. Hard cuts on the beat; every cut changes both shot size and shot type. Lyrics (the lead says “hello” in each language in turn, with precise lip-sync): English “Hello”, Chinese “你好”, Japanese “こんにちは”, Korean “안녕하세요”, Portuguese “Olá”, Thai “สวัสดี”, Spanish “Hola”, Arabic “مرحبا”. Shot 1 [0:00–0:03] Low-angle extreme wide establishing shot; Steadicam pushes in slowly through golden light and sea mist. Line 1 (English). Hard cut. Shot 2 [0:03–0:05] Close-up of the lead rapping to camera, handheld whip-in; sea sparkle out of focus behind. Line 2 (Chinese). Hard cut. Shot 3 [0:05–0:08] Macro insert, locked off: a guitarist’s fingers strumming fast; sand and salty mist drift past the lens. Line 3 (Japanese). Hard cut. Shot 4 [0:08–0:10] Three-quarter medium shot of a band member, slow creeping orbit; metal hardware catches the low sun. Line 4 (Korean). Hard cut. Shot 5 [0:10–0:13] A band member at the shoreline, fast lateral dolly; he turns to camera as a wave breaks behind. Line 5 (Portuguese). Hard cut. Shot 6 [0:13–0:15] The drummer at the water’s edge, fast handheld tilt up, wind and spray in his hair. Line 6 (Thai). Hard cut. Shot 7 [0:15–0:18] Tight aggressive push-in on the lead mid-flow; band silhouetted against the dusky sea. Line 7 (Spanish). Hard cut. Shot 8 [0:18–0:20] Heroic extreme wide of the full band, aggressive handheld push; they step toward camera on the beat as a wave crashes and golden light flares behind them. Line 8 (Arabic). 4000K white balance, teal-and-amber grade, 35mm, shallow depth of field, film grain, sea haze, golden-hour glow. Solid, premium texture. Rhythmic rap performance, precise lip-sync, nodding on the beat. No subtitles, no text overlays, no dissolves, no duplicated people — hard cuts only. Total length 20 seconds.

Inputs
Image 1
Output
480p draft → 1080p final

Draft mode: preview first, then render

Generate a quick low-resolution draft to check the scene layout, camera moves, action and whether the model understood your prompt. When you are happy, turn that draft into the final high-quality video.

  • Drafts are always 480p and cost the same as a normal 480p video.
  • The final video is rendered at 1080p and reuses the draft’s prompt, files, seed, audio setting, ratio and length, so the key elements stay consistent.
  • A draft can be turned into a final video for 14 days after it is created.
Prompt

The girl holds a fox, opens her eyes and looks gently at the camera; the fox hugs her in a friendly way; the camera slowly pulls back; the wind moves her hair and you can hear it.

Inputs
Step 1 · Draft video
Step 2 · Final video

Also supported

Text to video

Generate a video from a text prompt alone.

Web search

The model can decide to look up fresh information — products, weather and so on — to make the video more current.

02

Output settings: resolution, ratio, length and format

Four settings control what Seedance 3.0 delivers. Here is what each option does and when to use it.

Resolution

  • 480p — 8-bit, fastest; also used for drafts.
  • 720p — 8-bit, the default.
  • 1080p — Full HD video with clear, detailed visuals.

Duration

  • Choose any whole number from 4 to 30 seconds, or Auto (the default), which lets the model pick a length that fits.
  • Video editing always uses Auto: the result matches the source length (it may be a fraction of a second shorter), and the source must be 4–30 seconds long.

Aspect ratio

Choose 16:9, 4:3, 1:1, 3:4, 9:16 or 21:9, or leave it on Auto (the default):

  • Text-to-video and reference-to-video: the model picks the best of the six ratios for your prompt.
  • Video editing and extension: the output keeps the ratio of the video being edited or extended.
  • First/last frame: the output follows the first frame. These three tasks only support Auto.
Output size in pixels for each resolution and aspect ratio
Resolution16:94:31:13:49:1621:9
480p854×480752×560640×640560×752480×854992×432
720p1280×7201112×834960×960834×1112720×12801470×630
1080p1920×10801664×12481440×14401248×16641080×19202206×946

Output format

  • MP4 (default) — standard color, plays everywhere: web, phones and social platforms.
  • MOV — high color precision (H.264, 4:4:4 chroma, PCM audio) for grading, keying and compositing. Recommended as both input and output for editing and extension. Play it with IINA, VLC, mpv or ffplay.

03

How to write a Seedance 3.0 prompt

Treat Seedance 3.0 as your film crew and write like a director: say what each file is for, sum the shot up in one line, script it beat by beat, then add the details that hold the whole video together.

The basic structure

  1. Step 1Map your filesWhen you upload references, number them in upload order (Image 1, Video 1, Audio 1) and say what each one is for: who is the character, whose voice, which motion, which location.
  2. Step 2Sum it up in one lineSubject + location + event + genre or style + any special camera move.
  3. Step 3Script it beat by beatSplit the story with timestamps or “Shot N” and describe each part: what we see, camera, action, dialogue and sound. Prefer positive descriptions.
  4. Step 4Close with global detailsAdd what stays true throughout: camera position and movement, setting, sound and mood.

Example: panda cub on a grassy slope

Prompt

Realistic nature-documentary style, cinematic natural light, a warm afternoon on a grassy forest slope: a round panda cub rolls down the hill. The panda’s black-and-white fur is fluffy and real; it is small, chubby and clumsily cute. The slope is green forest ground covered with grass, moss, clover, soil, pebbles, twigs and a few little yellow flowers, with tall tree trunks blurred in the background. Low-angle medium-wide shot with a slight handheld feel, mostly steady, always keeping the panda in frame. 0s–3s: The cub lies on the slope, round as a ball, and starts rolling slowly sideways, clumsily pressing the grass flat. A breeze blows; sunlight from the top left filters through the trees in dappled patches. 3s–8s: The cub rolls to the lower right and comes to a stop, turning from its side onto its belly. Its round face turns toward the camera, front paws on the grass; it settles into a comfortable position in the foreground grass, lifts its head a little and lowers it again, making soft little grunts. Low angle, slight handheld feel, gently following the panda toward the lower right. Natural depth of field: foreground blades softly blurred, the panda sharp, the forest behind softly out of focus. Natural ambience — wind and the soft thuds of the panda rolling. Warm, real and natural overall.

Output

Three quick rules

  • Follow the order: subject + action/event + scene and setting + visual style + camera moves and cuts + sound. Leave out whatever you don’t need.
  • Give every file a job: refer to files as @Image 1, @Video 1, @Audio 1 and say exactly what each one provides (look, motion, voice) — and what it doesn’t.
  • Mark sound with symbols: music in ( ), sound effects in < >, dialogue in { } and subtitles in 【 】. For non-English lines, name the language before the line.

Referencing many files

The more files you upload, the more the mapping matters. Number files in upload order and tie each one to a name in the text. Don’t rely on labels written inside the images — writing “Tom” on a picture and then “Tom walks into school” in the prompt can confuse or duplicate characters.

  • List every subject’s mapping one by one; use a list when there are many people.
    • “the swordsman from Image 1”
    • “Images 1–2 are Person 1, using Audio 1; Images 3–4 are Person 2, using Audio 2”
    • “Image 1 is the lead, Tom, speaking with the voice from Audio 1”
  • Say exactly what to take from each file. If you only want part of it, say which part.
    • “the spell-casting motion from Video 1 and the orbiting camera from Video 2”
    • “the lighting and color grade of Image 1”
  • When a reference is already precise, just point to it instead of re-describing it.
    • “Follow the motion and camera of Video 1 exactly, in the same order.” — no need to spell out “raises a hand, turns, the camera slowly orbits…”

Editing prompts

Say exactly what to change and how far. Add timestamps for partial edits, and describe the change as going from A to B.

  • Examples:
    • “Only edit the man’s line in Video 1: change it to ‘Stay away from me!’ with a Boston accent.”
    • “In Video 1, change the man drinking coffee at 4–6s to mopping the floor; keep everything else the same.”
    • “Edit task: replace the woman on the right of Video 1 with the woman from Image 1.”

First and last frames

There are two ways to pin the opening and closing shots:

  • Upload the images in the first-frame and last-frame slots (recommended). The output locks to the first frame’s aspect ratio and matches it exactly.
  • Upload them as normal references and name them in the prompt. The ratio is not locked; the result stays close to the images but may not match them exactly.
    • “Image 1 is the first frame.”
    • “Image 3 is the first frame and Image 5 is the last frame.”

Timestamps

Timestamps (in whole seconds) keep the story in order. If a time range has little in it, the model improvises; if it has too much, you get rushed cuts or skipped beats — so budget time sensibly. Don’t use timestamps for frequency (“shakes head 3 times in one second”).

  • Continuous time ranges — no gaps like “0–3s… 5–6s”:
    • “0–3s… 3–7s… 7–15s”
    • “[1s–4s]… [4s–8s]… [8s–12s]”
  • A specific moment:
    • “At 5s, a fast whip to the left as a transition”
    • “At 2s, a bolt of golden lightning strikes down from the top of the frame…”
  • Relative timing:
    • “Tom stands frozen; 3 seconds later everyone around him shakes their heads”
    • “…after the shutter clicks, the frame freezes for 1 second”

Negative instructions

You can switch off subtitles and individual layers of sound — effects, background music and dialogue.

  • Subtitles:
    • “Don’t add subtitles for the dialogue.”
    • “No subtitles.”
  • Audio:
    • “No BGM — only ambient and action sounds.”
    • “No sound at all.”

Camera language

Common film terms work as-is: shot sizes (extreme wide, wide, medium, close-up, extreme close-up), moves (push in, pull out, pan, truck, follow, orbit, dive, tilt up, handheld shake) and angles (low angle, top-down, first-person). Popular techniques such as one continuous take, dolly zoom, aerial, FPV, bullet time and speed ramps work too.

  • Turn niche jargon into “term + plain description”.
    • “Rack focus: the focus shifts smoothly — the trees in the foreground blur while the person in the background comes into focus.”
  • For transitions, state when and how they happen.
    • “At 5s, a fast whip to the left (wipe left + natural dissolve).”

Actions and expressions

  • Actions: summarize (“does a few sets of high knees and flips”, “the two fight up close”) and only spell out the one or two moves you want people to remember. Don’t repeat the same action.
  • Expressions: describe what the face does instead of using idioms.
    • Instead of “eating with relish”, write “smiling with satisfaction, taking big bites”.

Untextured 3D Model Reference & Rendering

An untextured 3D model video shows models without surface textures. The models can be simple or detailed; use their camera movement, animation and composition as references for a finished shot.

  • Say what to take from the untextured 3D model video.
    • No lighting in the untextured 3D model animation, only camera and motion: “Follow the camera and motion of Video 1…”
    • Lighting changes you want to keep: “Follow the lighting changes of Video 1 and…”
  • If you add reference images, map them to the untextured 3D models.
    • “The man in grey from Image 1 is the red model in Video 1; the red-haired girl from Video 2 replaces green model 2 in Video 1.”
  • Still describe the video you want in detail, and make sure it matches the untextured 3D model scene. For shapes without a reference image, describe their look.

Storyboard grids

A grid image of storyboard panels gives the model your shot plan.

  • Keep it under about 15 panels. Very large grids (e.g. 18 panels) often freeze or shuffle the order, and panels limit the model’s freedom, so make sure each one is accurate.
  • Avoid messy, over-sharpened AI storyboards and panels crowded with text.
  • Keep the prompt consistent with itself and with sensible camera and motion design.
  • The video won’t follow the grid exactly — the model improvises a bit. For a strict match, use keyframes instead.
  • Recommended: simple stick-figure or line-art panels plus a structured prompt — (1) map the files, (2) give the story summary, (3) describe every shot in full, filling in what the panels don’t show; timestamps help.
  • For concept boards or keyframe designs, a short prompt is enough.
    • “Follow the storyboard in order to build a complete story with smooth, logical camera work.”
Storyboards to avoid

Line-art storyboard example

Prompt

Files: storyboard @Image 1, bedroom @Image 2, Leo @Image 3, Mia @Image 4, the book “Happy Days” @Image 5. Shot 1: [Wide, locked off, eye level, rule of thirds] A snowy winter night. By the floor-to-ceiling window, the man stands side-on with his hands in his pockets, watching the snow; the girl stands beside him, quietly looking at him. Calm and restrained; snow keeps landing on the glass. Shot 2: [Medium over-the-shoulder] The girl’s back in the foreground; the man turns and looks at her warmly; she lowers her head in silence; snow keeps falling outside. Shot 3: [Medium close-up, diagonal composition] The man slowly holds out the book “Happy Days”; the girl reaches up and takes it. Shot 4: [Close-up on the girl, centered] She hugs the book to her chest; her eyes redden and a tear slowly rolls down. Sad. Shot 5: [Close-up on the man, angled] He smiles softly, quietly watching her cry, a trace of wistfulness in his eyes. Shot 6: [Wide, locked off] The girl turns and slowly walks out of frame; the man stands alone at the window, hands in pockets, watching the snowstorm. The room is empty and silent.

Concept storyboard

Keyframe references

When the video must follow your boards exactly, upload each panel as its own reference, in order, and say so in the first sentence: “Use Images 1 to 7, in order, as keyframes.”

Prompt

Use Images 1 to 7, in order, as keyframes. Among mountains above a sea of clouds, a blue-and-pink long-tailed spirit fish glides through the air; the camera slowly pushes toward an old town built into the mountainside and settles on the pagoda at the summit. The scene turns into an elegant Chinese-style hall: the spirit fish flies in through a window and lands in the round pool at the center, swimming lazily. Finally the view cuts to a dim old temple, where a white-bearded monk stands with his back to us, gazing at a huge framed painting — the very hall and the spirit fish in its pool. New Chinese ukiyo-e illustration style.

Image 1
Image 2
Image 3
Image 4
Image 5
Image 6
Image 7

What’s different from Seedance 2.0

Timestamps work

Seedance 2.0 only followed shot numbers; Seedance 3.0 follows whole-second timestamps.

Multi-view references

Using several views of one subject as a reference wasn’t recommended in 2.0; 3.0 supports it.

Any aspect ratio

2.0 had six fixed ratios. 3.0 can output any ratio between 0.4 and 2.5 by following your input files.

Better quality for video-to-video

MOV output keeps color, brightness and audio-video sync more consistent in extension and editing.

04

Seedance 3.0 prompt examples

Subject, motion, audio and style references work the same way as in Seedance 2.0. These examples show the reference modes new in 3.0.

Simple untextured 3D models → animated short

Feed a untextured 3D model animation that carries the cuts, camera moves, paths and lighting, then add character, scene and prop images to control the look.

  • Simple models work best: build people, objects and animals from basic shapes.
  • Keep only the body in the untextured 3D model animation and avoid limbs or wings. If they are there, animate them fully, or they will look stiff in the render.
Prompt

Use the untextured 3D model video (Video 1) as the only reference for camera moves, pacing, shot sizes, the subject’s path and staging for the whole film. Keep its shot order, camera changes, movement and rhythm exactly — no restructured, added or reordered shots, no change to how the subject moves. Using the keyframe images for each stage, make a 30-second cinematic 3D animated short: dreamy, fairy-tale, warm, full of childhood fantasy. Characters must match the keyframes for each stage; don’t change their design. Expressions change with each scene. 0–3s (first frame: Image 2): From a wide top-down view the camera slowly pushes toward a little girl sitting on the bedroom rug, playing with a toy plane. She stands, turns left and throws the plane into the air with her right hand. 3–5s (Image 3): The plane flies left to right past the hanging star ornaments, and the girl rides it into an imaginary sky. 5–8s (Image 4): The camera keeps tracking from the side, orbiting around the girl. 8–10s (Image 5): The camera orbits behind the plane she is flying; it starts a slow dive toward the sea. 10–19s (Images 6 and 7): The girl swims down into the deep; a manta ray glides into frame. 19–23s (Image 8): She bursts out of a rift in time into a fantasy universe. 23–24s (Image 9): The girl and a planet in the foreground flip forward and dissolve away. 24–28s (Image 9): The top-down camera keeps pushing in; the girl is asleep on the rug. 28–30s (Image 10): The camera pushes into the picture book, then Dad walks into frame. Overall: every shot follows its keyframe; the untextured 3D model video is only a reference for camera, movement and character animation, not for what things look like. Output a 30-second 16:9 widescreen video.

Inputs
Video 1
Image 2
Image 3
Image 4
Image 5
Image 6
Image 7
Image 8
Image 9
Image 10
Output
Side-by-side comparison

Detailed untextured 3D models → re-rendered scene

Use complete, detailed 3D models without surface textures as the basis for a richer, more polished render.

  • Use a clean, complete untextured 3D model video with no path lines, axes or camera cones — they tend to leak into the result.
Prompt

Render the untextured 3D models in Video 1. No BGM — only ambient and action sounds. Render: a cyberpunk city at night in deep blue and purple, dense skyscrapers with huge holographic billboards and neon between them, a few aircraft weaving through the sky with faint lights and a soft mechanical hum. The character is a small raccoon in a black stealth suit, seen as a silhouette, stepping carefully. It moves across the roof of one of the skyscrapers.

Inputs
Video 1
Output
Detailed untextured 3D model preview

Storyboard grid reference

A nine-panel storyboard sets the shot structure; extra images lock the setting and two characters for a 30-second, nine-shot film.

Prompt

Image 1: nine-panel storyboard, for the overall shot structure, shot sizes and camera rhythm. Image 2: live-action reference of a rocket launch site on a prairie at dusk, for composition, warm gold sunset against cool twilight blue, and a realistic color photographic look. Image 3: look of Subject 1 (the guardian robot). Image 4: look of Subject 2 (the old woman). [Subjects] Subject 1 (guardian robot): based on Image 3 — a weathered retro robot from the near future, faded teal metal body with patches of rust, a domed head, two glowing red round eyes, a thin antenna, long slender jointed limbs; tall, about twice a person’s height. Subject 2 (old woman): based on Image 4 — a small, frail elderly woman, silver hair in a low bun, deep wrinkles, a bright golden floor-length dress with blue-and-gold embroidery on the chest, reluctant to let go; she only reaches the robot’s chest. Setting (prairie launch site at dusk): based on Image 2 — a near-future prairie at dusk, the sky fading from warm gold to cool blue; a launch pad on the horizon holds a white rocket, steam rising; knee-high grass ripples in the wind, vast and empty. [Style] Live-action color feature film, photoreal throughout; color 35mm film texture with fine natural grain, rich cinematic grade, IMAX large-format feel. Handheld with a breathing sway, shallow depth of field, grass, sparks and ash drifting in the foreground, a slight Dutch angle; warm golden sunset, cool dusk blue and explosive warm orange clash hard. 16:9 landscape. Near-future, tender disaster-film mood: quiet, tragic, protective, reluctant to part. [Exclude] Black and white, monochrome, grayscale, desaturated; hand-drawn, sketch, line art, illustration, comics, animation; storyboards or sketches; tilt-shift miniature, toy look, plastic CG, greasy overexposed CG. [Shots] (9 shots ≈ 30s) Shot 1 (0–3s): Extreme wide, low angle looking up — the two seen from behind on the dusk prairie, gazing at the distant rocket. Robot: “I’m right here. I won’t let go.” Shot 2 (3–6s): Frontal medium shot, handheld — the robot supports the old woman. Shot 3 (6–10s): Close-up — her eyes full of reluctance. Old woman: “Fly safe, my child. Come back to me.” Shot 4 (10–14s): Extreme wide, tilting up — the rocket lifts off trailing thick white smoke. Old woman: “There he goes… there he goes.” Shot 5 (14–18s): Extreme wide — the rocket explodes and breaks apart in mid-air. Old woman: “No… no, no—” Shot 6 (18–22s): Extreme close-up — her pupils contract, a tear slides down. Old woman: “…he was almost there.” Shot 7 (22–25s): Close-up widening to a medium close-up — she breaks down sobbing. Old woman: “Bring him back! Please — bring him back!” Shot 8 (25–28s): Ultra-low angle looking almost straight up — the robot wraps around her like a protective dome. Robot: “Don’t look up. I’ve got you.” Shot 9 (28–30s): Extreme wide from behind — the two hold each other tight, in silhouette. Robot: “I’m still here. I’ll stay… as long as you need.”

Inputs
Image 1
Image 2
Image 3
Image 4
Output

Keyframe reference

Six keyframes become one continuous vertical pixel-art video with 8-bit music.

Prompt

Using @Image 1 to @Image 6, make a vertical, one-take pixel-art martial-arts video with Chinese-style 8-bit martial-arts background music. A uniform light-blue background throughout, consistent pixel style, clean and clear. Shot 1: Hold still on the ink-brush “Jianghu Saga” logo from @Image 1 on the light-blue background for about 1 second. Shot 2: The text area from @Image 1 disappears; the pixel close-up of the male martial artist’s face from @Image 2 slides up from the bottom, blinks, looks at the camera, then quickly drops out of frame. Where the logo was, the blue pixel “Secret Martial Arts Manual” book from @Image 3 appears. Shot 3: Right after, the little pixel hero from @Image 4 leaps up from the bottom of the frame and bumps the blue diamond question block above; dark-blue bold text “Into the Jianghu today!” pops out above it. He lands in the pose from @Image 4, waves, gets ready to run, then turns and runs to the right (running pose from @Image 5). The camera follows him right as he jumps out of the right edge of the frame. Shot 4: The UI screen from @Image 6 slides in from the right; the hero jumps in from the top right, lands at the lower right of the big “27th Day of the Third Month” text, and throws his arms wide in an enthusiastic pose, freezing on that frame. Consistent pixel martial-arts art style on a light-blue base. Smooth, continuous camera that reads as one take with continuous movement and following; natural transitions between elements and fluid character motion, no stutter or flicker; text and UI crisp and stable.

Inputs
Image 1
Image 2
Image 3
Image 4
Image 5
Image 6
Output

05

Video editing with Seedance 3.0

Add, remove or change what’s in an existing video with a prompt — optionally guided by reference images — or edit its audio and dialogue.

Edit with an instruction

Change the content of a video with a prompt alone.

Prompt

Keep the framing, camera position, lighting and acting rhythm of @Video 1. Only rewrite the woman’s looks and expression: let her age naturally from her twenties to sixty; the restraint in her eyes slowly melts, tears glisten and slide from the corners of her eyes, the corners of her mouth lift little by little, and finally she smiles through her tears. One continuous take, no jump cuts or flicker; her features stay put as she ages.

Inputs
Video 1
Output

Edit with reference images

Add images to steer the edit — here a new setting and new costumes for two fighters, with the choreography unchanged.

Prompt

Turn the plain two-person fight video @Video 1 into a tense bare-handed sizing-up before a duel with blades. Replace the setting with a medieval stone castle platform, an old courtyard, a terrace outside a mountain fortress or a simple stone-paved dueling ground — castle walls, wind, mist and distant mountain ridges behind, flat stone underfoot @Image 1. Dress the man in dark clothes as in @Image 2 and the man in light clothes as in @Image 3. Keep the moves and the original rhythm unchanged. AI effects only enhance the environment and texture: wind in the clothes, light mist, a little dust at contact points, cold metallic reflections, slight grain and an epic grade. Restrained, realistic, classic hard-hitting duel mood. Background music cut to the beat.

Inputs
Image 1
Image 2
Image 3
Video 1
Output

Edit the audio

Translate the dialogue into another language and re-sync the lips.

Prompt

Translate the spoken dialogue in the video into Chinese, with no subtitles; adjust the lip movements precisely to match; keep everything else the same.

Inputs
Video 1
Output

06

More creative workflows

Extend a clip, turn a photo set into a video, or bridge two videos with a seamless transition.

Video extension

Continue an existing video. The volume may shift slightly from the input; extending a video that Seedance 3.0 generated itself gives the smoothest join. Choose MOV as the output format.

Prompt

Continue @Video 1 for 5 more seconds: a bee flies in and lands on the painting; then an extreme macro close-up of its legs and belly covered in golden pollen grains. The bee beats its wings and takes off; the camera follows it to another flower of the same kind. In slow motion, pollen shakes loose from the bee’s fuzz and falls right onto the stamens — the moment of pollination, magnified.

Inputs
Video 1
Output
Stitched result

Watch around the 15-second mark — there’s no visible seam.

One-click montage from photos

Turn a set of photos into a finished, edited video with music.

Prompt

Turn all the images into one finished video, in any order: a hand-drawn, animated-doodle, cut-out style café vlog following a puppy in different cute outfits posing for photos at a café. Add playful, internet-style audio or BGM. The images can move slightly, like live photos, but don’t change them — keep them highly faithful to the originals.

Inputs
Image 1
Image 2
Image 3
Image 4
Image 5
Image 6
Image 7
Image 8
Output

Seamless transition between videos

Generate a bridge that joins two clips without a visible cut.

Prompt

Join Video 1 and Video 2. The view in Video 1 rushes up to the top, quickly turns back and dives straight down, then transitions seamlessly and naturally into Video 2; during the transition the mahjong tiles slowly turn into tall buildings and the whole scene changes to match. Don’t change either of the uploaded videos.

Inputs
Video 1
Video 2
Output

07

Seedance 3.0 FAQ and troubleshooting

Common problems with generation, editing and translation, and how to fix them in your prompt.

Why does Seedance 3.0 look different from Seedance 2.0 with the same prompt?

What happens

The two models have different visual tastes. The same reference image and prompt can produce clearly different styles, so mixing their clips can look inconsistent.

How to fix it
  1. To keep the Seedance 2.0 look, feed the 2.0 video into a Seedance 3.0 extension task and choose MOV as the output format. The original style carries over far better.
Seedance 2.0 output
Seedance 3.0 output, same inputs
2.0 video extended with MOV output — style preserved
Reference image used
Seedance 2.0 still frame
Seedance 3.0 still frame
Three-frame comparison

Why do characters’ eyes glow?

What happens

The model can over-react to strong emotions and make the pupils glow blue, red or another unnatural color.

How to fix it
  1. State it with top priority: the characters have normal human eyes that don’t glow, and ambient light only touches the face outline.
  2. Don’t describe emotions as eye effects. Swap intense words like “fanatical” or “utterly shocked” for milder ones like “amazed”.
✕ Before — Intense emotion words make the pupils glow
✓ After — Normal-eyes constraint + milder words
Glowing pupils in the “before” video
Before (shot 2)
Prompt

…Gao Yan says: “It goes straight into the blood? That’s magic!” His tone is full of fanatical worship of modern medicine.

After (shot 2)
Prompt

…Gao Yan says: “It goes straight into the blood? That’s magic!” His tone is full of amazement at modern medicine (the character’s eyes do not glow).

How do I stop on-screen text from being misspelled?

What happens

Letters in words get swapped, added or dropped (E often becomes A or I), and Chinese characters can be replaced by look-alikes.

How to fix it
  1. Split the word into single letters and ask for them to appear one after another, so the model doesn’t guess or simplify.
  2. Put the exact text into an image and upload it as a reference.
✕ Before — “DIHETAO” misspelled as “DIHERTAO”
✓ After — Letters appear one by one — spelled correctly
Before
Prompt

Create an opening title sequence for a beetle game called 'DIHETAO'. The text should be white colored. There should be many beetles.

After
Prompt

Create an opening animation for a beetle-themed game. The text should be white, and the scene must feature a large number of beetles. Then, the letters "D", "I", "H", "E", "T", "A", and "O" should appear in sequence.

Why do subtitles appear when I asked for none?

What happens

Even with “no subtitles” in the prompt, captions can still show up. It happens far less than in Seedance 2.0, but it can’t be ruled out completely.

How to fix it
  1. Don’t repeat words from a line afterwards and pile tone, expression or action notes onto them — that invites subtitles.
  2. To direct each line differently, use the format “Character line (emotion): text”.
  3. Make sure your reference videos have no subtitles of their own, or the instruction may be ignored.
✕ Before — Repeated words with tone notes → subtitles appear
✓ After — “Line (emotion): text” format → no subtitles
✕ Before — Reference video with built-in subtitles → output copies them
✓ After — Subtitle-free reference video → clean output
Before
Prompt

Granny Zhou says, in an old, hoarse voice: {Like… so alike. You finally came, the village is saved.} After “like” a short pause; “so alike” in disbelief; “you finally came” with a trembling voice; “the village is saved” full of relieved excitement.

After
Prompt

Granny Zhou line (in disbelief): “Like…” Granny Zhou line (voice trembling): “So alike.” Granny Zhou line (overjoyed): “You finally came, the village is saved.”

Why is my JPG image rejected with a decode error?

What happens

Some JPGs fail with an error such as “invalid image: heic: decode failed”. The file isn’t corrupt — it uses an unusual, asymmetric chroma subsampling that falls outside the standard 4:2:0, 4:2:2 and 4:4:4 formats.

How to fix it
  1. Re-save the image as a standard JPG (4:2:0) in any image editor or converter, or convert it to PNG, then upload it again.

Why is my edited video slightly shorter than the original?

What happens

Editing output can be about 0.33 seconds shorter than the source video.

How to fix it
  1. Switch to reference-to-video mode and set the length. It isn’t a strict edit, so content may differ a little, but the duration is exact.
  2. If you stay with editing, trim the input to 8n+1 frames (n is a whole number). The output then matches the source length.
Input video
✕ Before — Edit output — about 0.33s shorter
✓ After — Reference mode — exact length
Input trimmed to 161 frames (6s 17f)
✓ After — Edit output — also 6s 17f
Checking timecode in an editing app
Timecode detail
Edit prompt used
Prompt

This is a video editing task. Only change what the instruction asks for; keep the subject, framing, timing and style the same. Instruction: remove the red lines, red numbers and red arrows from the video.

Why are results unstable when one prompt asks for too much?

What happens

When a single generation has to handle several references and edits at once, the result becomes unreliable — here, a red path line leaks into the video.

How to fix it
  1. Break the job into simple tasks, generate or edit them one at a time, and combine them in your editor.
✕ Before — Everything in one request — the red path line shows
✓ After — Split into a reference task and an edit task — clean
Step 1 input: image with the flight path drawn on
Step 1 output: the hat flying along the path
Step 2 input: original video, with Step 1’s output as a reference
Step 2 output: the hat lands on his head
Original single prompt
Prompt

Edit Video 1: add a hat that flies along the path marked in Image 1 onto the man’s head; don’t show the red path.

Step 1 · reference task
Prompt

Generate a 5-second video: a hat flies along the path marked in Image 1 onto the man’s head; don’t show the red path.

Step 2 · edit task
Prompt

Edit Video 1: add a hat flying onto the man’s head; the hat’s look and path follow the hat in Video 2; keep everything else the same.

Why does changing the aspect ratio in an edit go wrong?

What happens

When an edit changes the frame shape, the model invents extra picture to fill the new area, and the original content or timing drifts.

How to fix it
  1. Use reference-to-video instead, and add a concrete second-by-second description of what happens, stressing frame-by-frame, shot-by-shot timing and complete picture content.
Input video
✕ Before — Edit output — content and timing drift
✓ After — Reference mode with a scene timeline — consistent
What to add
Prompt

Timeline: [00:00–00:01] Extreme close-up of a tiger’s face staring into the lens, a double exposure of shimmering waves over its outline, warm golden-orange fur in the low sun. [00:01–00:04] The tiger walks along the beach toward the camera, low angle, rolling light-blue waves and a sunset-yellow sky behind. [00:04–00:06] A white seabird soars against a bright blue sky; the camera tracks its flight.

Why do fingerprint-like patterns appear in grass or leaves?

What happens

When a high-resolution AI-generated image is used as a reference, dense textures such as grass and leaves can show fingerprint or footprint-like patterns.

How to fix it
  1. Resize the input image so it is no larger than the output resolution. It happens less at 1080p than at 720p. This will keep improving in future versions.
✕ Before — High-res reference → fingerprint texture
✓ After — Image resized to the output resolution → gone
Original high-res reference
Fingerprint texture close-up
Resized reference

Why is there background music when I asked for none?

What happens

Even with “no background music” in the prompt, some sound effects can turn into a music bed.

How to fix it
  1. In the prompt: say the soundtrack contains only the characters’ voices and no musical track at all, and list every synonym in one go — music, background music, BGM, score, instrumental, melody, synth, ambient pad — so the model can’t slip around it.
  2. Write the audio rule at both the beginning and the end of the prompt.
  3. After generation: separate the vocals and remove everything else. Note that this removes sound effects too.
✕ Before — “No BGM” stated, but effects become music
✓ After — Synonyms listed + rule repeated at start and end
✓ After — Vocals separated in post (effects removed too)
Audio rule to add at the start
Prompt

Audio policy (applies to ALL shots): the soundtrack contains ONLY the specified spoken dialogue plus faint natural night ambience at a low, constant level. Absolute silence otherwise. No music, no BGM, no score, no instrumental, no melody, no soundtrack, no synth, no ambient pad, no drone, no swelling tone, no crescendo, no musical build or stinger at any moment.

Repeat after the shots
Prompt

Audio: ONLY faint crickets and light wind at a low constant level, unchanged from the previous shots. Absolute silence otherwise. No music, no score, no swelling tone, no rising pad, no synth, no ambient drone, no crescendo.

Why does audio translation leave some of the original language?

What happens

After translating a multi-character video, bits of the original language remain, voices land on the wrong character, or the lips don’t match the new language.

How to fix it
  1. Say that every voice must be translated — including background lines, whispers, exclamations and filler sounds.
  2. Give each character’s full translated lines with a timeline.
  3. Ask for word-by-word lip-sync and spell out the acceptance criteria: no trace of the original language, no subtitles, lips in sync, and nothing else in the picture or sound changed.
Original multi-character video
✕ Before — Leftover original language, voices swapped
✓ After — Full lines and timeline per character — correct
Before
Prompt

Translate the voices in the video into Chinese and change the lip movements precisely to match; keep everything else the same. No Korean or Chinese subtitles.

After (condensed)
Prompt

Translate ALL spoken lines in @Video 1 — male and female, including background dialogue, whispers, exclamations and filler words — into Mandarin Chinese, with precise lip-sync and no subtitles. Keep each speaker’s order, timing, emotion, pauses and volume; never swap male and female voices. Change nothing except the spoken language and lip movements. Chinese lines and timeline: 00:05–00:09 Woman: “So for dinner tonight, want to get your favorite fried chicken?” 00:16–00:21 Man: “More than fried chicken… I love times like this, having coffee with you.” 00:23–00:24 Woman: “Oh…” Final check: every line in Mandarin, zero Korean audio or text, no subtitles, lips in sync, everything else unchanged.

Why is there bubbling or echo noise in the audio?

What happens

The audio can contain unexpected bubbling, echoes or a faint whistle.

How to fix it
  1. Remove descriptions of water, waves and echoes from the prompt to make this less likely. It will keep improving in future versions.
✕ Before — Bubbling noise in the audio
✓ After — Water, wave and echo wording removed — less noise

Ready to direct your first scene?

Copy any prompt from this guide, swap in your own files and generate it with Seedance 3.0.

Start creating