Seedance 2.0 Prompt Guide

Extract selected elements from your assets (such as a subject, style, scene, or sound) and generate a brand-new video.

• Best for: motion transfer, subject reuse, mood borrowing, and similar tasks
• Recommended patterns:

• Image reference: Referencing the <Subject N> in <Image N>, generate...
• Video reference: Referencing the <motion / camera work / style / sound> in <Video N>, generate...
• Audio reference: Referencing the voice timbre in <Audio N>, generate...

Make local or global modifications on top of an existing video. Anything you do not mention stays unchanged by default.

• Best for: local replacement, subject removal, attribute changes, and similar tasks
• Recommended patterns:

• Add an element: clearly describe the <element traits> + <when it appears> + <where it appears>
• Modify an element: Strictly edit <Video N>, changing the <original trait> to the <new trait>
• Delete an element: name the element to remove; for elements that should stay unchanged, explicitly calling them out in the prompt improves results

Continue the original video along the time axis, keeping the audio-visual style, subjects, and narrative consistent.

• Best for: continuing a storyline, extending an action, filling in missing segments
• Recommended patterns:

• Extend a video: Extend <Video N> forward/backward, generating...
• Track completion: <Video 1> + <transition description> + then <Video 2> + <transition description> + then <Video 3>

The three task types above can also be combined.

• Best for: referencing one asset while editing another
• Recommended pattern: Referencing the [reference dimension] of <Image/Video N>, strictly edit <Video X>, [specific edits]

Slogans

Prompt template:

「text content」+「when it appears」+「where it appears」+「how it enters」,「text traits (color, style)」

Reference Materials

Result

Prompt

Hand-drawn comic style. Three people sit together eating the fried chicken from Image 1, in a warm, friendly mood. The frame then gradually blurs, and the text “快乐尽在 Seedance” (“Joy is all in Seedance”) appears in the middle of the frame.

Subtitles

Prompt template:

Subtitles appear at the bottom of the frame, reading “…”, perfectly synced to the audio.

Result

Reference Materials

Prompt

Generate a video with a voiceover. A deep, calm male voice says: “In the vastness of the cosmos, our world is but a fleeting moment. And yet, within it, life flourishes against all odds.” The scene transitions slowly from night to dawn — stars fading as the sun rises from behind the mountains. Subtitles appear at the bottom of the frame following the narration.

Result

Reference Materials

Prompt

The two people in the image chat in an office. The woman speaks first: “你每次卡点到，是不是很享受这种刚刚好的感觉？” (“You always arrive right on the dot — do you enjoy cutting it that close?”) The man replies with a smile: “我有我的节奏” (“I have my own rhythm”). The conversation feels casual and natural, with matching dialogue subtitles at the bottom of the frame.

Speech Bubbles

Prompt template:

「Character」says: “…”; as they speak, a bubble appears next to them with the line written inside.

Result

Reference Materials

Prompt

The two people in Image 1, wearing sportswear, run on the school track. The girl looks at the boy and says with a confident smile: “We can definitely do it!” Cut to a close-up of the boy, who answers hesitantly: “Are you sure?” Cut back to a medium close-up of the girl, who says brightly: “Yes!” — upbeat and firm. A bubble appears next to whoever is speaking, with the matching line inside.

Result

Reference Materials

Prompt

Referencing the girl in Image 1 and Image 2: in a strawberry field, she picks a strawberry, takes a bite, and says with a smile: “This is the real deal!” A bubble appears around her with the line written inside.

Seedance 2.0 supports multi-view subject references as well as multi-image references such as scene images and storyboards.

If the order of images matters, upload them in order and refer to them precisely in the prompt as Image 1, Image 2, … Image n.

Multi-View Subject Reference

Just make the reference target clear — the model responds to instructions including (but not limited to) these examples. Products:

Consumer electronics

Result

Reference Materials

Prompt

Extract the camera from Image 1, Image 2, and Image 3, and swap the background to white. The camera sits on a white table; the lens focuses on it in close-up, then rotates slowly around it as the subject, clearly showing its front, sides, and back.

Home goods

Result

Reference Materials

Prompt

Against a warm-toned home scene, a medium shot presents the thermos from the reference images. The camera pushes in smoothly to a close-up; a hand enters the frame naturally, grips the body of the thermos and lifts it, with the camera following the hand's slight rotating motion as it shows the product.

Characters:

Reference Materials

Result

Prompt

Referencing the woman in Image 1, Image 2, and Image 3, generate a scene of her eating cake in a coffee shop.

Multi-Image Reference

Logo reference

Result

Reference Materials

Prompt

The backdrop is a neon-lit aerial walkway in a future city, aircraft weaving between holographic ads. Referencing the girl in Image 2: first a medium shot of her releasing a silver hover-lantern carrying a holographic projection, then the camera pulls back to reveal a sky full of floating lanterns. The frame gradually blurs, and the logo from Image 1 appears. Overall style: 3D cyberpunk sci-fi animation.

Multi-subject reference

Result

Reference Materials

Prompt

Referencing the cat and dog in the images: in a cozy apartment, the dog lies eating from its bowl; the cat walks over and taps the dog with a paw; the dog stops eating when it sees the cat, and the cat nestles against the dog. Warm color palette.

Multi-element reference

Result

Reference Materials

Prompt

The scene is set inside the restaurant from Image 4, bustling with customers. The girl from Image 1, dressed in the outfit from Image 2, is tidying items on the counter. The boy from Image 3 is a customer; he walks up, hoping to ask the girl for her contact info. The logo from Image 5 stays visible in the bottom-right corner of the frame throughout.

Multi-panel storyboard reference

Result

Reference Materials

Prompt

Referencing the storyboard in the image, generate an intense fight scene. Each panel's composition appears in order, after which the two fight fiercely.

Storyboard reference

Result

Reference Materials

Prompt

Referencing the framing of panel Image 3: a girl waits for her dad to finish cooking and says: “아빠， 배고파요！ 밥 다 됐어요？” — her look references Image 1. The camera then pans right, switching to the framing of Image 4; the dad, whose look references Image 2, answers: “거의 다 됐어， 조금만 기다려！” Cut back to a close-up of the daughter's slightly disappointed face as she says: “아직 멀었어요？ 맛있는 냄새 나는데…” Then cut to a close-up of the dad, who says: “이제 진짜 금방이야。 ＂빨리빨리＂ 하지 말고 손부터 씻고 와！”

Seedance 2.0 supports video reference — just be explicit about what to generate and what to reference.

If the order of videos matters, upload them in order and refer to them precisely in the prompt as Video 1, Video 2, … Video n.

Action Reference

Film & TV

Result

Reference Materials

Prompt

Referencing the character movement and shot language of Video 1, generate a fight scene between Image 2 and Image 1 — Image 2 is the person on the left, Image 1 the person on the right. Intense background music.

Marketing

Result

Reference Materials

Prompt

Referencing the horse's galloping form in Video 1, generate a golden steed galloping across grassland, then freeze its magnificent running pose as it turns into a horse-shaped gold pendant.

Camera Movement Reference

Result

Reference Materials

Prompt

Referencing the camera movement of Video 1, make a concept video of a tech park: the high-rise in Image 1 is the visual center, same first-person dive, conveying the futuristic feel of the campus in Image 1.

Effects Reference

Film & TV

Result

Reference Materials

Prompt

Referencing the golden particle effect in Video 1: as the character in Image 2 plays the flute, the same particle effect swirls around them.

Novelty effects

Result

Reference Materials

Prompt

Referencing the effect in Video 1, give the girl in Image 1 the same wings, with an identical wing-growing trajectory.

Seedance 2.0 supports video editing: adding, deleting, or modifying elements; extending a video forward or backward; and track completion.

If the order of videos matters, upload them in order and refer to them precisely in the prompt as Video 1, Video 2, … Video n.

Add / Delete / Modify Elements

Add elements

Result

Reference Materials

Prompt

Add fried chicken, pizza and other snacks to the countertop in Video 1.

Delete elements

Result

Reference Materials

Prompt

Clear the other parts and tools off the desk in Video 1, keeping the desk clean and tidy — only what the two of them hold in their hands remains.

Modify elements

Result

Reference Materials

Prompt

Replace the perfume in Video 1 with the face cream from Image 1, keeping the motion and camera work unchanged.

Video Extension

Track Completion

A single reference image often contains multiple subjects. To reference a specific object precisely, you must define your subjects explicitly. A subject can be a person, a prop, or a scene.

The model's internal representation decouples space and time. So for a complex video, the ideal prompt shape is a timeline of shots: split the video into a few shots and describe each one dynamically, in the order events happen — who + where + doing what + how the camera moves.

Practical advice: write a simple “Shot 1 / Shot 2 / Shot 3” breakdown for the video, then merge it into the full prompt.

• Weak example: “A man runs nervously through the streets; the frame feels very cinematic.”
• Strong example:

• Shot 1: Side view of a street alley; the man slowly starts to run, with a sense of urgent breathing.
• Shot 2: The man crashes into a fruit stall; the camera whips quickly and lands on a close-up of his terrified face.
• Shot 3: The man vaults over a low wall and disappears; the camera slowly pulls back and holds on the empty street.

Specific rules

• Use markers like Shot 1, Shot 2, Shot 3 and organize content in the order events happen (primary first). Don't force a duration for each segment — let the model pace the story naturally.

Each shot is best organized in this order:

1. Camera movement or transition: e.g. “wide shot slowly pushing in”, “locked-off camera”, “cut to…”.
2. Subject action and expression: the key actions and expression changes of the main character / object.
3. Position or spatial change: the scene, position, or spatial relationships of the subject.
4. Audio information: the sound effects, dialogue, or background music for that shot.

• Specific body parts + quantified intensity — Describe actions down to hands, legs, head, shoulders and back, and add amplitude, speed, and force. Examples: slowly raises a hand, quickly turns the head, kicks off the ground hard, lowers the head slightly.
• Prefer slow, gentle, continuous small movements — Favor slow, soft, connected micro-actions; avoid sprinting, big jumps, violent rolls and other high-burst, large-dynamic moves. Examples: walking slowly, gently raising a hand, slightly lowering the head, sitting down in one smooth motion.
• Describe transitions between actions — Spell out the momentum and hand-off between consecutive actions so movement stays coherent. Examples: uses the momentum of the turn to raise a hand, transitions naturally from stillness into raising a hand.
• Externalize emotion into physical detail — Express emotion through concrete body detail instead of abstract words like “very sad” or “furious”. See the table below:

The model understands camera vocabulary very well — just use standard cinematography terms directly, e.g. “medium shot, close-up, wide shot, slow push-in, smooth lateral tracking, locked-off camera”.

Quality, style, and constraint terms are key to controlling the output. They define the creative boundary for the model, unify image quality and art direction, and suppress visual defects and random drift — a necessary setup for stable, compliant results.

1. Image quality — defines clarity, detail texture, and lighting feel, raising the baseline quality of the output. Example: high definition, rich detail, cinematic texture, natural color, soft lighting.

2. Style — sets the overall art direction and visual tone, unifying the artistic atmosphere. Example: cyberpunk cold blue-purple palette, vintage film, fresh Japanese-style tones.

3. Constraints — extremely important. Constraints effectively suppress visual defects, deformities, and unreasonable elements, bounding generation and improving stability. Common templates:

• Avoid subtitles: “keep the video free of subtitles”, “avoid generating any text or subtitles”
• Avoid logos: “do not generate logos”
• Avoid watermarks: “do not generate watermarks”

Two examples showing how the advanced formula and its elements come together in a real prompt.

Seedance 2.0 can generate common on-screen text. The model automatically matches a suitable style and color to the context, and also lets you specify the text's color, style, entrance animation, timing, and position in the prompt. Prefer common characters and avoid rare characters and special symbols for the best rendering. Supported scenarios currently include slogans, subtitles, and speech bubbles — see the Text Generation examples above for concrete patterns.

• Continuous long take (video extension): best for dialogue-driven “quiet” scenes inside a single location — long conversations, emotional build-up, movement along a single path — for an immersive one-take feel.
• Scene / action turns (segmented stitching): best for plot turns or complex, fast “action” scenes — chases, fights, montages. Generate segments independently and edit them together to preserve pacing and visual impact.

Real projects usually combine both: extend to get a coherent conversation first, then stitch in cutaways or transition shots — keeping immersion while varying the rhythm.

Think of your assets as playing four functional roles:

1. Character anchor: locks the character's appearance
2. Scene tone-setter: locks the environment and style
3. Camera reference: locks the shot language and action rhythm
4. Rhythm & mood: uses audio to control emotion and voice timbre

Recommended setup (4–5 assets total): 1–2 character images (face close-up / full body) + 1 scene image + 1 camera-reference video + 1 audio clip.

Keep dialogue in a single language. Avoid mixing languages in the same line (proper nouns excepted).

Using symbols consistently helps the model tell different kinds of information apart:

Symptom — The generated character doesn't match the reference image, or the face changes mid-video (ID drift), sometimes tripping review filters for resembling celebrities.

Root cause — The face reference isn't effective enough:

• Mixed reference images: the face reference is merged into one image together with full-body pose shots, outfit references, or detail shots.
• Face too small in frame: in a mixed reference image the face occupies too small a share of the image, so the model under-weights facial features and gets distracted by background or other elements.

Solution — Strengthen the independence and weight of the face reference:

1. Prepare a dedicated face close-up: besides the full-body shot, add an image containing only the character's head (a headshot — face only, neutral expression is best, minimal shoulders/background).
2. Define the subject clearly in the prompt: “Subject 1's facial features reference Image 1 (headshot); makeup and styling reference Image 2 (full-body shot)”.
3. Put important assets first: the more precisely an asset needs to be referenced, the earlier it should appear in the prompt.

Symptom — The prompt never asked for subtitles, but the generated video contains them.

1. Add explicit constraints to the prompt: “keep the video free of subtitles”, “avoid generating any text or subtitles”.
2. If text in your reference images/videos isn't essential, remove it first (for example with Seedream / Seedance image or video editing), then use the text-free asset as input.
3. Where your use case allows, generate in landscape first (subtitles appear noticeably less often than in portrait), then crop to portrait in an editor.

Symptom — The prompt never mentioned watermarks, but the generated video contains logos or watermarks from other video platforms.

Solution — Add explicit constraints to the prompt: “do not generate watermarks”, “do not generate logos”.

Symptom — You want 2D or 3D anime output, but the reference image is fairly photorealistic and the prompt doesn't stress the style — so the result can drift into a live-action look.

Solution — Add explicit style constraints such as “2D Japanese anime style” or “3D Chinese-style animation”. For tighter control, convert the reference image into the target style first, then generate the video from it.

Symptom — After generating an extension and joining it to the original, the seam can show a visible jump or a step backwards in content.

Solution — For now, patch it in post by aligning keyframes (a root fix will come with model iteration):

1. Import the clips to be joined into CapCut or another editor.
2. At the first seam, trim 6 frames off the end of the earlier clip.
3. Also trim 1 frame off the start of the later clip.
4. Repeat for every join point.
5. Export and check that the joined video plays smoothly.

Symptom — In scenes with many characters, especially when a character turnaround (three-view) sheet is used as reference, the same frame can contain two identical copies of one character.

Root cause:

• Character subjects aren't clearly defined in the prompt, so the model can't precisely tell the roles apart.
• Turnaround / multi-view sheets confuse character recognition, producing duplicated look-alike characters.

1. Bind each character to its reference: define every character clearly and note which image it maps to, keeping the format consistent. Example: Zhang San (from Image 1) throws the green passbook at the standing Li Si (from Image 2).
2. Add a global constraint at the end of the prompt: “Throughout the video, never show characters with identical appearance, clothing, and accessories; no clones or twin effects; keep only a single instance of each character per frame; no duplicated characters.”
3. Optimize reference assets: prefer single-person photos; avoid three-view / multi-view sheets.
4. Trim the prompt: don't paste a full script as the prompt — redundant copy confuses the model. Cut irrelevant text and keep instructions focused.

Symptom — Using a generated video as input for further extension degrades image quality. Repeated extensions stack the degradation, with mottled color patches appearing especially on faces.

Solution — Mitigations for now (a root fix will come with model iteration):

1. Convert the original video to a white clay-render with Seedance 2.0 first, then use that as the extension input. Reference prompt: “Convert the video into a white 3D model: all characters become pure-white 3D models with no color, no texture, no shadows, pure white background, stable structure, fluid motion.”
2. Prefer high-resolution images as reference assets.
3. Limit how many times you chain extensions; avoid stacking many rounds.

Symptom — Describing a specific visual effect in text can miss. Example: the prompt asks for “the number 2999 entering as a countdown animation”, but the generated digits scroll with chaotic jumps instead of a proper countdown.

Solution — Define the effect with a reference video: feed a clip of the target effect so the model precisely understands its form and motion logic. Example: “The number 2999 enters the way it does in Video 1.”

Symptom — With more than 4 reference characters, output stability drops: the video may show the wrong number of people (missing or extra) or duplicated characters.

Solution — Mitigations for now (a root fix will come with model iteration):

1. Generate group images in steps: split the characters into groups of at most 4 per image. For example, 6 characters can be split into 2 groups of 3, each generating one image.
2. Then image-to-video: use those grouped images as reference assets to generate the final video.

Symptom — When using a reference audio clip to set the voice, the generated video's voice can deviate noticeably from the reference timbre.

Solution:

1. Add detailed voice-timbre descriptions to the prompt, e.g. “speaks in the low, warm, slightly gravelly middle-aged male voice of @Audio 1”.
2. Keep the dialogue's tone and phrasing style close to the reference audio — it improves timbre fidelity and stability.