Seedance 2 Prompt Guide

A useful Seedance 2 prompt tells the model which reference to use, what should happen, and how the camera should capture it. Start with one clear idea, assign each uploaded reference a job, and add only the details that matter to your scene.

This guide follows the Seedance 2 prompting approach used by Video Agent. It also includes examples for the AI video generator, where you can supply additional reference inputs. Choose Seedance 2 for these examples; available references and settings differ between models and versions.

Start with one image and one scene

For a first test, open the AI video generator, select Seedance 2, choose Text/Image to Video, and upload one character image. Use it as the reference for a 10-second video.

@image1 as character anchor. Keep the same face, hairstyle, and outfit as the reference.

[0-5s] Medium shot with a slow push-in: the character stands on a quiet rooftop at dusk, looks toward the horizon, and takes a gentle breath. Soft blue light, reflective mood.
[5-10s] Close-up, camera locked off: the character turns toward the camera with a subtle smile. A light breeze moves their hair, warm city lights softly blurred behind them.

Here, @image1 identifies the reference image. The next sentence states what to preserve. The two timed lines describe motion, camera behavior, and atmosphere. This example needs no reference audio.

Set the actual video duration to 10 seconds too. Writing “10 seconds” in the prompt does not replace the duration setting.

In Agent, select a scene image and edit the Scene Context field. Adapt this example's timestamps to the duration shown on that scene card.

Where should you write the prompt?

For example, an Agent project idea can be plain language:

Create a reflective music video about a singer returning to their hometown. Use quiet rooftops and empty streets at dusk, soft blue and amber light, and restrained camera movement. Keep the protagonist's face, short dark hair, and black jacket consistent throughout.

You can give Agent this direction before planning individual shots. You do not need to assign @ tags to the whole project first.

Use the Translate to English button

In the AI video generator, Translate to English also refines the prompt for the selected model. Select your model, input mode, duration, and reference files first so the translation uses the right context.

The translation is instructed to:

• Organize your existing details into clear camera, action, mood, and scene instructions.
• Use the selected model's reference syntax: Seedance references such as @image1, Wan reference names where supported, or natural-language guidance for other modes.
• Preserve your intent, valid reference tags, existing timestamps, and shot order. Keep deliberately quoted dialogue, lyrics, and requested on-screen text in their original language.
• Avoid adding characters, props, actions, unavailable references, or unwanted lip-sync. A simple shot should stay simple.

The result replaces the text in the prompt field and remains editable. Review the reference roles, duration, and creative meaning before generating. If you change models or reference files, check the prompt again. Translation does not upload files or generate the video.

What do @image1, @audio1, and @video1 mean?

These tags refer to the reference files supplied to this particular video generation. Each media type is numbered separately, starting at 1.

Numbering within each media type follows the official reference-order convention. Official Seedance tutorial

Two images and one audio clip are referenced as @image1, @image2, and @audio1. The audio does not become @audio3.

• Keep the tag intact: write @image1, not @image 1 or @图片1.
• You do not need to rename your file to image1. The tag follows the reference order, not the filename.
• Upload or select the reference first. Typing a tag does not attach a file.
• Recheck the order after adding, removing, or replacing references.
• Use only tags for files included in the current generation. A file elsewhere in your project is not automatically available to the model.

References in Agent and the video generator

Agent scene generation uses the selected scene image as @image1. It does not automatically submit every character image uploaded to the project. Having two images in your project does not mean you can use @image2 in every scene.

For a prompt that combines two images or references a video, use the AI video generator with Seedance 2 and attach those references there. To upload a reference video for @video1, choose Reference to Video. Assign roles according to the order of the uploaded references. The multi-image and video examples below assume you are using that input setup.

A song uploaded to Agent is not always a reference audio clip for video generation. Generating from an Agent scene card attempts to use a trimmed audio segment; if trimming fails, that generation can proceed without it. The background automatic pipeline currently generates without reference audio and adds the project music when merging. For a controlled lip-sync test, use the video generator with a character image and the intended audio segment attached.

In the video generator, choose Audio → Reference Audio to attach that segment. Generate Audio controls generated sound; it does not attach a reference audio file.

Give every reference a clear job

Put reference instructions at the beginning, before describing the shot. Explain what the model should take from each file.

as character anchor means “use this image to keep the character's appearance consistent.” It does not make the character perfectly identical in every output; review the face, hair, and clothing after generation.

Be specific about the role of a second image. If both images show the same person, say so. If they show different people, identify each person and describe who performs each action. If a video should guide only the camera, say that instead of also asking the model to copy its characters and setting.

Build the shot in a useful order

After the reference declarations, write:

Camera movement → subject and action → emotion → setting and lighting → optional visual texture.

Instead of “a beautiful cinematic singer,” describe something visible:

@image1 as character anchor.
Slow push-in from a medium shot: the singer stands beside a rain-streaked window, lifts their gaze toward the camera, and gently rests one hand on their chest. Their expression shifts from hesitation to resolve. Soft side lighting, muted blue tones, shallow depth of field.

Choose one main camera instruction for each shot. “Camera locked off” and “fast handheld orbit” ask for incompatible movement if applied to the same moment.

Use timestamps to plan a short sequence

The Seedance 2 Agent workflow plans 4–15-second clips. A first test of 8–12 seconds is a useful starting point. Short clips can stay on one shot; longer ones can use two or three sub-shots.

For a 12-second clip:

@image1 as character anchor. Keep the same outfit throughout.

[0-4s] Wide shot, slow push-in: the character stands beneath a streetlight on a rain-wet street. Quiet, expectant mood.
[4-8s] Medium shot, gentle lateral tracking: the character takes two steps forward and looks up. Warm light catches their face.
[8-12s] Close-up, camera locked off: the character holds their gaze with a faint smile. The street behind them falls softly out of focus.

• Start at 0 for each generated clip, even if that scene occurs later in the song.
• Put each timed shot on its own line.
• Cover the full selected duration without gaps, overlaps, or an end time beyond the clip length.
• Following Agent's shot-planning guidance, allow at least 4 seconds per sub-shot. Do not force several cuts into a 4-second clip.
• Timestamps communicate intended pacing. They do not guarantee frame-exact editing; check and trim the result when needed.

For standalone clips, try “Shot 1, Shot 2…”; exact timing is unreliable. Official Seedance prompt guide. Agent timestamps are planning cues.

Copyable examples

Use only an example whose listed references are actually attached. Adjust the selected duration to match its timestamps.

Character plus background — 12 seconds

Inputs: two images in the video generator: a character portrait first, then a background image. No audio reference is needed.

@image1 as character anchor. @image2 as background scene. Preserve the character's face and outfit from @image1 and the location's architecture from @image2.

[0-4s] Wide shot, slow push-in: the character stands in the location shown in @image2, taking in the surroundings. Soft evening light, quiet atmosphere.
[4-8s] Medium shot, gentle tracking: the character walks a few steps through the scene and turns toward the camera. Natural posture and relaxed hands.
[8-12s] Close-up, camera locked off: the character pauses with a thoughtful expression. Keep the same lighting and background style.

With only one image, remove the @image2 instructions and describe the background in words.

Singing performance with reference audio — 10 seconds

Inputs: one character image and a matching 10-second vocal segment supplied as reference audio. This example is for a generation that actually receives both files.

@image1 as character anchor. Lip-sync with @audio1. Clear mouth movement while singing. Keep the same face, hairstyle, and outfit.

[0-5s] Medium shot, slow push-in: the singer faces the camera and sings with a calm, heartfelt expression. Mouth clearly visible, shoulders relaxed, soft warm side lighting.
[5-10s] Close-up, camera locked off: the singer's expression grows more emotional with the vocal phrase. Subtle eyebrow movement, natural breathing, hands outside the frame.

For a first lip-sync test, use clear isolated vocals with little or no background music, keep the reference segment within 15 seconds, and match it to the intended shot. A visible mouth, a moderate vocal pace, and restrained motion make the result easier to evaluate. The prompt requests synchronization; check the generated mouth movement against the audio.

Speaking to camera — 8 seconds

Inputs: one character image and an 8-second spoken audio segment.

@image1 as speaking character. Lip-sync with @audio1. Clear mouth movement while speaking.

[0-4s] Medium shot, camera locked off: the character faces the camera and speaks naturally. Attentive expression, relaxed shoulders, simple studio background, soft even lighting.
[4-8s] Cut to a close-up, camera locked off: the character emphasizes the final point with a subtle nod and a small smile. Keep the mouth visible and the same background and lighting.

Use the spoken audio to supply the words. Writing dialogue in the prompt alone does not supply an @audio1 reference.

Dance with a camera reference — 8 seconds

Inputs: one character image and one video attached as a camera reference in the Seedance 2 video generator. No lip-sync instruction is needed.

@image1 as character anchor. @video1 as camera movement reference only. Keep the character and outfit from @image1; use a neon-lit studio as the setting.

[0-4s] Smooth lateral tracking inspired by @video1: the dancer performs a short side-step sequence with relaxed, controlled arm movements. Blue and pink studio lighting, energetic mood.
[4-8s] Gentle orbit inspired by @video1: the same dancer turns and finishes in a balanced pose. Maintain the same outfit and lighting, full body visible.

Choose a reference video with the type of camera movement you describe. If you want to copy choreography instead, explicitly assign the video an action reference role and describe the camera separately.

Keep a character consistent across scenes

1. Use consistent character references or scene images that already preserve the same appearance.
2. Repeat the character instruction at the start of each video prompt when that reference is supplied.
3. Repeat visible details that matter: hairstyle, jacket, accessories, and color palette.
4. Change the action or emotion deliberately while keeping identity details stable.

@image1 as character anchor. Keep the same short dark hair, black jacket, and silver necklace. Maintain the same character appearance across this sequence.

Saying “the same person as the previous scene” is not enough by itself: each generation needs the relevant visual reference. For a longer music video, plan several clips around musical phrases, review each one, and then merge them. A prompt alone does not extend a single Seedance 2 generation beyond its supported duration.

Troubleshooting

Change one important detail at a time so you can tell which revision helped.

Before you generate

• Confirm the selected model and the references supported by that input mode.
• Check that every @ tag points to a supplied file, with the correct number and role.
• Match the selected duration to the final timestamp; use clip-local times starting at zero.
• Keep image prompts about a still frame and video prompts about movement over time.
• Check that camera, action, and reference instructions agree with one another.
• For lip-sync, supply the intended audio segment and keep the mouth visible.
• Generate and review one scene before applying the approach to a full music video.

Start a Video Agent project · Try a Seedance 2 clip · Read the Video Agent guide