Seedance Prompt Guide: A Free Template and 4 Realistic Examples

Picture the job: a vertical first frame of two friends in a laundromat, and a two-line joke. What goes in the prompt box? Not a paragraph of adjectives. Seedance responds best to a short brief, like the one a director hands a camera crew: what the clip is, which references to follow and what must not change, the beats in order with every spoken line quoted, one camera behavior, the sound, and a few closing rules such as no subtitles.

That method carries over between the two versions in Spicy Studio's Video Lab, Seedance 2.0 and Seedance 2.5. Below: a free fill-in template based on ByteDance's published prompt guides, four example prompts, the settings and credit costs, and fixes for clips that look fake. No account needed to read it.

Want a tested long-form prompt? The sketch workflow in the Spicy Studio playbook (Spicy Pro) includes the full Seedance 2.0 prompt behind its example street interview. The prompt runs to about 650 words, in labeled parts from format and identity to camera rules, audio and realism.

ByteDance documents both models on its BytePlus developer site, in the Seedance 2.5 prompt guide and the Seedance 2.0 prompt guide. Here is what they say, as of October 2026, that matters most for realistic talking clips.

Summary first (2.5). Open with one sentence covering subject, location, event, style and camera movement, then the detailed plot, then notes on what must stay consistent, such as camera angle, setting, sound and mood.

Who, where, how, then constraints (2.0). The 2.0 formula goes from who is doing what, to where and in what mood, to how it's shot, and ends with style, image quality and constraint words.

Number every reference. Name each upload by its order (Image 1, Video 1, Audio 1) and tie it to its job in the text, such as the voice from Audio 1. A name written inside an image isn't enough.

Different timing syntax. Seedance 2.5 understands timestamps in whole seconds (0–3 s, 3–7 s), with no gaps between ranges. Seedance 2.0 ignores timestamps and follows shot numbers (Shot 1, Shot 2); ByteDance warns that forcing exact durations on it can give odd results.

One camera move per shot. The 2.0 guide asks for a single movement type per shot, because stacking a push, a pull and a pan makes the picture unstable.

Mostly positive wording. The 2.5 guide prefers saying what you want and keeps negatives for subtitles and audio. The 2.0 guide treats short constraint lines (no subtitles, no logo, no watermark) as essential.

One language. The 2.0 guide says not to mix languages in dialogue, apart from proper nouns. ByteDance says 2.5 supports native generation in more than 10 languages.

No length target. Neither guide sets a word count. The 2.0 guide asks for concise wording without repetition or contradictions and advises against pasting a whole script in as the prompt.

Here's that advice as seven labeled lines. Copy it, replace every bracket and keep each line short. Type the labels or drop them; the order is what counts.

Summary: [Vertical 9:16] [handheld phone video] of [who, with an adult age] [doing what] at [place], [time of day], [length] seconds, one continuous take.

References: Start from the provided first frame. Keep [name A] and [name B] exactly as they appear in it: same face, hair, clothes and build.

Beats: On 2.5, time ranges: 0–[x] s: [action]. [Name A] says: “[line 1]” [x]–[y] s: [reaction]. [Name B] says: “[line 2]” [y]–[end] s: [last reaction]. On 2.0, the same beats in order, as one take or as Shot 1 and Shot 2.

Camera: [one behavior: handheld at chest height with slight shake, or locked off on a tripod].

Sound: [the voices] plus [two or three sounds the place would really make]. [No background music.]

Keep throughout: [light source], real skin texture, everyday phone image quality, faces and outfits unchanged.

Rules: No subtitles. No captions. No on-screen text. No watermark. No logo.

Name each person (Maya, Theo) and reuse the name exactly; the 2.0 guide recommends one stable label per character. Describe the background through things that move, like a door or traffic, so the scene doesn't freeze. Spicy Studio's Seedance panel has no negative-prompt box, so constraints go in the Rules line. And cast a fictional character or yourself: a real person's face belongs in a prompt only with their consent.

Each example uses the template, trimmed to what the clip needs.

1. Two-person sketch with timestamps (Seedance 2.5, 10 seconds).

Prompt: Vertical 9:16 handheld phone video of two friends in their late twenties at a laundromat at night, 10 seconds, one continuous take. Start from the provided first frame and keep Maya and Theo exactly as they appear in it. 0–3 s: Maya pulls a pink shirt out of the dryer and holds it up, frowning. 3–6 s: Maya says: “This was white an hour ago.” 6–9 s: Theo spots a red sock stuck to the sleeve and says: “Honestly? Pink suits you.” 9–10 s: Maya stares; he grins. Camera: handheld at chest height, slight shake, no zoom. Sound: their voices, tumbling dryers, a coin machine. No background music. Keep throughout: flat fluorescent light, real skin texture. No subtitles. No captions. No watermark.

Cost: 100 credits at 480p, 220 at 720p.

2. Walk-and-talk selfie (Seedance 2.0, 8 seconds).

Prompt: Vertical 9:16 selfie-style phone video, 8 seconds, one continuous take. Start from the provided first frame and keep the man exactly as he appears in it, beard, glasses and green rain jacket included. He walks along a wet city sidewalk, phone at arm's length, talking into the lens. First he sighs and says: “Moving tip: never trust anyone who says it's just a few boxes.” Then he points over his shoulder at a tall stack of boxes in the doorway behind him and raises his eyebrows. Camera: handheld selfie angle, natural arm sway, no cuts. Sound: his voice, light rain, a passing bus. No music. Keep throughout: overcast daylight, slight phone compression. No subtitles. No captions. No watermark.

Settings: Generate Audio on and 9:16 selected, since 2.0 defaults to audio off and 16:9. Cost: 56 credits at 480p, 120 at 720p.

3. Product demo with an AI presenter (Seedance 2.5, 8 seconds).

Prompt: Vertical 9:16 phone video of a woman in her forties at a kitchen counter showing a collapsible water bottle, 8 seconds, one continuous take. Start from the provided first frame and keep her and the bottle exactly as they appear, color and label included. 0–3 s: she holds the empty bottle at chest height and says: “[one feature from the brand's product page, such as: It folds flat when it's empty.]” 3–6 s: she presses it into a flat disc with one hand. 6–8 s: she sets it down and smiles at the lens. Camera: locked off at eye level. Sound: her voice, a quiet kitchen, the squeak of silicone. No music. Keep throughout: window light from the left, label sharp and readable. No subtitles. No captions. No watermark.

Ground rules for ads: build the first frame from a real photo of the actual product, script only what the brand states and never let the presenter pose as a customer; the FTC's rule on fake reviews covers reviews attributed to people who don't exist, AI-generated ones included. Label the clip as AI and disclose paid relationships. Some laws go further: New York requires advertisers to disclose synthetic performers. More in AI UGC Ads. Cost: 176 credits at 720p.

4. Silent reaction (Seedance 2.0, 5 seconds).

Prompt: The camera remains completely still at all times. Start from the provided first frame. A man in his thirties at a bakery counter watches the cashier hand the last croissant to the customer ahead of him. His shoulders drop, he glances at the empty tray, then looks up with a small, defeated smile. Ultra realistic. Natural movements.

Why it's short: with no dialogue, the animate workflow style beats the full template: plain actions in order, then its two-sentence realism closer. Turn on Camera Fixed. Cost: 35 credits at 480p, the cheapest Seedance clip.

The template runs on both versions because it describes a take, not a model. The settings panel is what changes.

Seedance 2.0: 5–15 seconds at 480p or 720p. Generate Audio starts off; turn it on for dialogue at no extra cost. There's a Camera Fixed toggle, and it takes up to 7 reference images, a 2–15 second reference video and up to 15 seconds of reference audio. It costs 7 credits a second at 480p and 15 at 720p; with a reference video, 4 or 9 a second over the reference and the output together.

Seedance 2.5: 4–30 seconds at 480p, 720p or 1080p, with audio on from the start. Up to 10 reference images, a 2–30 second reference video and up to 30 seconds of reference audio. Per second it's 10 credits at 480p, 22 at 720p and 39 at 1080p; with a reference video, 6, 13 or 23, again over reference plus output.

Frames or references, never both. On either version, a first or last frame switches the job to frame mode, and the reference images, video and audio are left out. Reference audio also needs a reference image or video beside it.

Don't fill every slot. For 2.0, ByteDance suggests four or five assets in all, such as one or two character images, a scene, a camera-movement clip and an audio clip.

Aspect ratio. Seedance 2.5 adopts the shape of your first frame. Seedance 2.0 defaults to 16:9, so pick 9:16 or adaptive yourself.

Which to use. 2.0 is cheaper for drafts and for sketches under 15 seconds; 2.5 suits longer takes, 1080p and timed beats. More on each: Seedance 2.0 on Spicy Studio and Seedance 2.5 features and pricing.

Seedance only moves what's already in your frame, and it won't repair a weak one. Get the still right and the prompt gets shorter.

Generate it. In the Image Lab, give Nano Banana 2 or GPT Image 2 your character's face reference plus a few labeled lines: scene, subject, who else is in frame and where, and style. Ask for a candid phone snapshot so it doesn't read as a render. Either costs 10 credits (Nano Banana 2 at 2K, GPT Image 2 at 1K medium).

Or face-replace a real photo. The first-frame workflow starts from a photo you took with the pose and setting you want, then swaps your character in from one face reference. Reuse that reference for every clip.

Enhance it. Run the frame through Nano Banana 2 or 2.1 with the playbook's default prompt, “Enhance the image quality,” or its longer version, which also asks the model not to change the person, pose, outfit or background.

Check it. Sharp faces, moving hands fully in frame, 9:16, and nothing in shot you don't want animated.

Copying a reference video? To borrow the movement, camera work and audio of footage you shot or are licensed to use, the character image goes in the references instead of the First Frame slot, with a much shorter prompt; the tested setup is in the complex motion workflow (Spicy Pro). For still-image prompting, see AI image prompt engineering.

With audio on, Seedance creates voices, mouth movement and room sound in one pass. These habits keep them in step.

Mark every line. Quote each line after the speaker's name, in speaking order. ByteDance's 2.0 guide also uses curly braces for dialogue and asks you to name any language other than English or Chinese, as in: says in Spanish {¿Qué pasó?}

Size the words to the clip. The playbook's sketch workflow aims for about 10 seconds of speech, 15 at most on 2.0. Too short cuts the punchline; too long and the dialogue wanders. Seedance 2.5 allows 30 seconds, but more speech means more drift, so add time only when the script needs it.

Which language. The playbook finds English sounds most natural, though other languages work. Don't switch mid-line.

Emotion once, before the line. The 2.5 guide notes that tagging single words with tone or action notes, or repeating a line's words, can trigger burned-in subtitles. Write it as Maya (deadpan): “This was white an hour ago.”

Captions in the edit. Keep the no-subtitles rule and add your own captions in your editor.

Need one exact voice? Seedance isn't the tool. Record the line and use Spicy Lip Sync, as shown in Make a Photo Talk With AI.

Most unconvincing clips trace back to a few causes. Find the symptom, change one thing and generate again.

It looks like a film, not a phone. Fix it at the source with a casual snapshot as the first frame. Then name the footage type (handheld phone video) and add texture such as natural light and slight compression.

The camera floats or swoops. Give it one behavior, plus what it shouldn't do: no zoom, no pan. On 2.0, use Camera Fixed.

Waxy skin. Ask for natural, textured skin and rule out beauty filters, in the image prompt and the video prompt.

Overacting. Swap emotion words such as very sad for small physical details: a half smile, a slow blink, raised eyebrows. ByteDance's 2.0 guide favors slow, small movements over sprints and jumps, and the 2.5 guide links intense emotion words to glowing eyes; calmer words and a normal human eyes line help.

Music sneaks in. ByteDance suggests stating that the audio holds only voices and ambience, listing what's excluded (music, background music, score, melody) and repeating that rule at the start and end of the prompt.

Subtitles appear anyway. Keep the rule, don't repeat a line's words elsewhere and use reference videos without text. The 2.0 guide adds that horizontal clips get unwanted subtitles far less often than vertical ones.

Face drift or twins. Tie each person to a stable name and image, use a clean headshot rather than a collage, and keep the cast to four or fewer; the 2.0 guide says stability drops beyond that.

A click at the end. Fade the audio out in your editor.

Then edit simply, the playbook way: captions on the first frame, a quick zoom on the punchline (it hides small glitches too) and no heavy effects.

You don't have to fill the brackets by hand. Give ChatGPT or another assistant the seven labels and a request like this: Write a Seedance prompt using Summary, References, Beats, Camera, Sound, Keep throughout and Rules for this sketch: [idea]. Characters: [names, adult ages, outfits, positions in the first frame]. Quote every spoken line, in order. Model: Seedance [2.0 or 2.5]; timestamps only for 2.5. Length: [10] seconds. One camera behavior. Then edit the answer: cut repetition and check that the beats fit the seconds.

Leave real people, celebrities, brands and copyrighted characters out of prompts and frames alike. Spicy Studio's Terms rule out realistic images of identifiable real people without consent, and TikTok's community guidelines bar using a private person's likeness without consent. When you post, use the platform's AI label; Do AI Videos Get Less Views? covers what that means for reach.

New accounts start with 25 free credits. That pays for a first frame and its enhance pass (10 + 10 = 20) but not a clip: the cheapest Seedance video, 5 seconds of Seedance 2.0 at 480p, costs 35.

Typical test runs: - 10-second Seedance 2.0 sketch at 480p, sound included: 70 credits - The same at 720p: 150 credits - 10 seconds of Seedance 2.5: 100 credits at 480p, 390 at 1080p - 10-second reference video plus 10 seconds of output, 2.0 at 480p: 80 credits

Small budget? The €2 Spark top-up adds 90 credits. With the free 25 that's 115, enough for a finished first frame (20) and one 10-second 2.0 sketch at 480p (70). A €5 top-up (240 credits) covers three such sketches, and the Basic plan is €10 for 600 credits a month (pricing). Top-ups never expire; plan credits reset each billing cycle. Draft at 480p and render only the keeper higher.

The full pipeline lives in the How to use playbook, free with an account (the sketch and complex motion workflows need Spicy Pro). The Free AI Video Course teaches it lesson by lesson, and 6 viral formats has more ideas to run through the template.