Prompt Seedance for image-to-video

Last updated: October 12, 2026

Seedance animates an approved still into a short clip that stays on-model, with sound generated in the same pass. Pick it when the look is already signed off and the job is motion: an idle, a spell cast, a living menu background.

Let the agent write it

You don't have to write any of this yourself. The Layer agent knows how to prompt each model it runs, from the same guidance this page is built on. Attach the still in chat, name the model and say what should happen, for example: Animate this with Seedance 2: a slow fire-spell cast, fixed camera, 4 seconds, with sound. The agent writes the prompt, shows you its plan, and runs it when you approve. Use the rest of this guide when you want to write or adjust the prompt yourself, and see Write a brief for the agent for how to brief it.

When to use it

Use Seedance when you have the frame and need it to move: character idles and actions, ability casts, ambient loops for menus and parallax scenes, and short trailer shots. It holds slow, continuous motion well and keeps the character's design, colours and style from the first frame.

It is weaker at fast, explosive action in one take (sprints, leaps, spins). Cut that into separate clips, or try another model from the model picker on the same frame. To generate a clip with no image at all, see Create a video.

Versions on Layer

Seedance 2

The version this guide's examples were made with. Clips run 4 to 15 seconds, with a first and a last frame, and native audio.

Seedance 2 Fast

The same inputs as Seedance 2, quicker and cheaper. Use it to find the motion, then render the keeper on Seedance 2.

Seedance 2.5

Clips up to 30 seconds, a Draft switch for a cheap low-resolution preview, and timed beats in the prompt (see below). The same prompt can look noticeably different on 2.5 and on 2, so keep one asset set on one version.

How to write the prompt

Camera, then the action step by step, then the style, then what must not appear:

Medium shot, fixed camera. The mage slowly raises the staff overhead with both hands as the ember swells into a ball of fire, then thrusts the staff forward and the fireball launches out of frame to the right; his cloak billows back with the force and he settles into a braced stance. Stylised painterly game art, colours unchanged. Keep it subtitle-free, no text, no watermark.

The first frame already fixes who is in the shot, where, and how it looks, so every word goes on what changes: what moves, in what order, how fast, and what the camera does. That prompt is the one in the clip above.

What works

Slow, step-by-step action

Write the action as a sequence, body part by body part, with its speed: "slowly raises the staff overhead with both hands, then thrusts it forward". We ran this against "explodes into action, leaps high into the air, spins and slams the staff into the ground" on the same frame and seed. The step-by-step prompt kept the mage on-model and landed every beat. The burst prompt chased a tiny, unrecognisable figure into the sky and ended in a full-frame explosion.

One camera move per shot

Pick one: fixed camera, slow push in, pull out, pan left, orbit, tracking shot, handheld. Several moves at once make the image unstable. For a rarer move, name it and describe it: "rack focus: the foreground blurs as the character comes into focus".

Motion, not description

Spend the prompt on what changes. In our test, restating the frame ("a young fire mage in a red hooded cloak stands on a mossy ledge...") before the same motion made no visible difference, so it is words you can drop.

JSON prompts

If you keep prompts as structured data, Seedance reads a JSON object as well as prose. We ran the spell cast above this way, on the same frame and seed:

{
  "shot": "medium shot",
  "camera": "fixed",
  "action": [
    "the mage slowly raises the staff overhead with both hands",
    "the ember swells into a ball of fire",
    "he thrusts the staff forward and the fireball launches out of frame to the right",
    "his cloak billows back with the force",
    "he settles into a braced stance"
  ],
  "style": "stylised painterly game art, colours unchanged",
  "constraints": ["no subtitles", "no text", "no watermark"]
}

Both versions landed every beat and kept the mage on-model. The JSON run held the camera a little steadier; the prose run's fireball read bigger. JSON is a way to organise a prompt, not a way to improve it: the model reads the keys as text, so none of them is a setting, and Duration, Generate audio and Aspect ratio still come from the form. ByteDance's own guides are written in prose. JSON earns its place when a script or workflow assembles prompts from parts, where named fields are easier to fill and to check.

Emotion as body language

Write "lowers his head, shoulders trembling" rather than "very sad". On 2.5, intense words such as "furious" can make eyes glow; use milder words, or add "normal eyes, no glowing eyes".

Name the style

End with the medium ("stylised painterly game art", "2D anime", "stylised 3D game cinematic") and "colours unchanged". Without it, stylised art can drift toward live action.

Several shots in one clip

On Seedance 2, write "Shot 1: ... Shot 2: ..." and let the model time them; it ignores timestamps. On Seedance 2.5, whole-second ranges work: "0-4s: the camera tracks alongside the mech as it steps forward. 4-8s: the camera holds as the thrusters ignite." Leave no gaps in the timeline.

Sound in the prompt

With Generate audio on, name the sounds you want ("a crackling fire and soft wind against the window") and put any spoken line in quotes. Say "no music" when you want none; on 2.5, if music still leaks in, repeat it at the start and the end of the prompt.

What to avoid

Expecting a negative prompt

Seedance has no negative prompt. Put exclusions in the prompt as plain constraints: "keep it subtitle-free, no text, no logo, no watermark". Portrait clips grow stray subtitles more often than landscape ones, so keep the constraint in every vertical clip.

Text and logos in the prompt

Generated lettering comes out misspelled. Paint the logo or title into the first or last frame, or add it afterwards in the video editor.

A frame in a different shape

A first frame whose shape differs from the output stretches or jumps. Set Aspect ratio to Auto so the clip follows the image, or crop the image to the shape you need first.

Counting on the last frame for a loop

Setting the same image as first and last frame brings the clip back to the same composition, but not to the same moment of the motion: in our test, the fire jumped at the seam as much as without a last frame. For a menu background, choose motion that hides a cut (flicker, drifting snow) and cross-fade the ends in the video editor.

Settings that matter

First frame and Last frame pin where the clip starts and ends. Duration runs from 4 seconds; find the motion at the shortest length, then render the length you need. Generate audio adds sound in the same pass. Aspect ratio sets the shape and resolution; Auto follows the first frame. Above the prompt, Camera, Staging, Look and Performance take reference sets for each.

Prompts to start from

Character idle

Locked-off camera. The mage breathes slowly, his cloak and the edge of his hood stirring in a light breeze, the ember on his staff flickering and shedding small sparks that drift upward. He turns his head slightly toward the valley, then back. Stylised painterly game art, colours unchanged. Keep it subtitle-free, no text, no watermark.

Spell cast

Medium shot, fixed camera. The mage slowly raises the staff overhead with both hands as the ember swells into a ball of fire, then thrusts the staff forward and the fireball launches out of frame to the right; his cloak billows back with the force and he settles into a braced stance. Stylised painterly game art, colours unchanged. Keep it subtitle-free, no text, no watermark.

Living menu background

Static wide shot, the camera does not move. The fire in the hearth crackles and flickers, the lantern flames sway gently, snow falls steadily past the window, and a faint haze of smoke drifts under the beams. The room stays empty, and the composition and lighting stay the same for the whole clip. Painterly fantasy game background. No characters appear. No subtitles, no text, no watermark. Sound: a crackling fire and soft wind against the window, no music.

Use it in Layer

  1. Add the approved still to a session's board.

  2. Click View on the image, then Turn into video. Create video opens in the side panel with the image as its First frame.

  3. Click the model name at the top of the panel and choose Seedance 2. The first frame stays in place.

  4. In the prompt box, write the camera, the motion step by step, the style and the constraints.

  5. Leave Duration short while you find the motion, and turn on Generate audio if the clip needs sound.

  6. Click Create video. The cost shows on the button before you click.

  7. When the clip lands on the board, click View to open it in the viewer and play it. Then change one thing in the prompt at a time and run it again.

Learn more

Related