Seedance

A good prompt is a specific work instruction, not a description that reads like ad copy. It states who is in which scene, what action they take, how the camera moves, and in what order events happen.

precise subject + action details + scene/environment + lighting & color tone
+ camera movement + visual style + image quality + constraints

First pin down "who" and "what they do," then describe "where" and "what the mood is," and then say "how to shoot it." Finish by refining the result with style, image quality, and constraints.

The video node has no separate field for a negative prompt. Write constraints, such as not adding a watermark or subtitles, as sentences within the prompt.

The basic syntax, such as defining subjects, notation symbols, and specifying reference materials, is the same in the Seedance 2.0 series and Seedance 2.5.

You can check the supported inputs, length, ratio, and resolution in Video models.

Defining subjects

A single reference image often contains several things, so to point to one of them precisely, define it as a subject first.

Define [key features] in <Image_N> as <Subject_N>

• Describe the subject with 2–3 features that do not change, such as clothing, hair, appearance, or category, so it can be recognized.
• When there are several people, define each one and use the same label consistently in the rest of the prompt.
• Image numbers follow the order in which you connect reference images to the video node.

Define the woman wearing a red dress and a straw hat in Image 1 as Subject 1

In a simple case where you do not define subjects in advance, tie each mention to its material with @.

girl @Image1 walks briskly to the dormitory entrance @Image2

Symbols for sound and subtitles

Wrap music, sound effects, dialogue, and subtitles in symbols to mark their type.

• If the dialogue is not in Chinese or English, state which language it is in.
• Use one dialogue language within a single prompt. Proper nouns are an exception.
• Music, sound effects, and dialogue are generated when sound is turned on in the video node. Connecting a reference video turns sound off, so turn it back on if you need it.

The model handles space and time separately, so the more complex the video, the better it works to write the prompt like a storyboard that follows the sequence of time. However, the 2.0 series is unreliable at precise timing, and forcing times in seconds can produce abnormal results.

Set only the order with Shot 1, Shot 2, and Shot 3, and let the model decide the length of each shot to suit the narrative. The overall length of the video follows the length you choose in the video node.

Write each shot in the following order.

Shot N: [Camera movement or transition] → [Subject action and expression] → [Position or spatial change] → [Audio]

Write only one camera movement per shot. Asking for a push, pull, pan, and tracking move all at once makes the frame unstable.

Calling materials by number applies when you connect reference images or reference videos. These two inputs cannot be connected together with start and end frames.

Divide materials by role into locking the character, setting the scene tone, and showing camera movement. The recommended setup is 1–2 character images (a close-up of the face and a full-body shot), 1 scene image, and 1 video showing the camera movement.

It is best not to connect as many materials as the maximum allows. You can connect up to 9 reference images in the 2.0 series and up to 30 in 2.5, but with many materials the model cannot judge which features to prioritize. Styles can then conflict with each other, the subject may not be recognized correctly, or the result may drift from what you intended.

The more precisely a material needs to be followed, the earlier in the prompt you specify it.

Reference <Subject_N> in <Image_N> to generate ...

Reference <Action/Camera movement/Style/Sound effect> in <Video_N> to generate ...

The prompt runs through the following four parts, in the order they are actually written.

1. Reference assignment — Use the girl in Image 1 as the main character and Image 2 as the dormitory scene, and refer to the camera movement in Video 1.
2. Subject definition — Define the girl wearing a navy cardigan in Image 1 as Subject 1.
3. Shot structure — Split the action and sound effects, dialogue, music, and chapter title across two shots.
4. Style, image quality, and constraints — Specify the style and color tone, a condition that keeps the face from deforming, and no subtitles or watermark.

Use the girl in Image 1 as the main character, use Image 2 as the dormitory scene reference, and refer to the camera movement in Video 1.

Define the girl in a navy cardigan in Image 1 as Subject 1.

Shot 1: At dusk, the camera follows Subject 1 in a medium shot as she walks to the dormitory entrance. She pauses at the doorway and takes a deep breath.
<footsteps echoing in an empty hallway>

Shot 2: The camera cuts to an indoor medium shot. Subject 1 enters and lowers her head, then looks up with a smile and says {I passed}.
(soft lo-fi piano plays quietly underneath)
【Chapter One: The Result】

Cinematic documentary style, warm tones, soft lighting. The character's face remains stable without deformation. Keep it subtitle-free. Do not generate a watermark.

The basic syntax, such as defining subjects, tying mentions to materials with @, notation symbols, camera terms, and specifying reference materials, is the same as in the 2.0 series.

2.5 generates up to 30 seconds at once, so a scene description sized for 5 seconds is not enough to control the whole video. Treat the model as a video content producer, and give the prompt a structure that follows the flow of visual storytelling.

For a 30-second video, it is best to include the following four items.

It is best to write the result you want as positive statements. Unlike the 2.0 series, 2.5 follows negative instructions only for subtitles and audio.

The 2.0 series sets the order with shot numbers, but 2.5 reads time ranges in seconds. Write ranges following these three rules.

1. Write them in whole seconds.
2. Leave no gaps between ranges. Do not skip ahead, as in writing 5-6s after 0-3s.
3. Put one clear action in each range. With too little content, the model fills the range in on its own; with too much, the cuts become overly fragmented or parts of the plot are dropped.

Klipse generates the video at the length you choose in the video node, so write the last range to end at that length. The example below assumes a length of 30 seconds.

0-5s: The chef sets an empty plate on the counter and picks up a spoon.
5-15s: Slow push in as sauce is drizzled across the plate in one line.
15-25s: A hand places three scallops down, one at a time.
25-30s: Pull back to a wide shot of the finished plate.

You can also specify a particular moment or a relative time.

Quick left sideways transition at the 5-second mark.

The frame freezes for 1 second after the main character presses the shutter.

Timing cannot be used to count how many times an action repeats. An instruction such as "nods three times in one second" is not reflected.

Because 2.5 accepts up to 30 reference images, it is best to also state how much to take from each material. Otherwise, backgrounds you did not ask for, passersby in the corner, and unwanted colors in a material can also appear in the result. Add only after the elements to take to limit the scope.

@Image 1 defines the woman's face and green jacket only.
@Image 2 defines the coffee shop counter and the window light only.
@Video 1 defines the camera move and the pacing only.

If a material shows exactly what you want, do not describe the same thing again. Writing only "strictly reference the action and camera movement in Video 1 and match the sequence" is enough; there is no need to describe details such as raising a hand or turning around again.

2.5 can reference images that show one subject from several angles together. However, if you only connect the images, it reads them as several different objects.

@Image 1 defines the front of the desk lamp. @Image 2 defines the left side of the same desk lamp. @Image 3 defines the back of the same desk lamp.
All three images show one lamp. The video must contain only one lamp.

The last sentence is what matters.

• Multi-angle images are best used when there are 5 or fewer subjects. With more subjects than that, using images from a single angle is more stable.
• When you need several angles, do not combine them into one image; connect a separate image for each angle.

Connect a storyboard as reference images, not as start and end frames. The result differs depending on whether you put several panels in one image or connect each panel separately.

To match the frames precisely, connect a separate image for each panel in order, and state in the first sentence of the prompt that the images are keyframes.

Use Images 1 to 7 in order as keyframes.

A storyboard combined in one image is referenced only for the broad flow of the plot.

• A storyboard of 15 panels or fewer, drawn as line art or with stick figures, works well.
• Do not put text on the storyboard.
• Use the prompt to fill in information the storyboard does not contain, such as action, camera, and style.

If the storyboard is at the level of a concept sketch, the prompt can be simple.

Construct a complete story plot according to the storyboard sequence, and use the shots in a reasonable and coherent way.

This reference method was added in Seedance 2.5. In a 3D tool, set up simple models, such as cylinders for people and spheres for objects, animate only the camera work and movement paths, and then upload that video to the video node as a reference video. Seedance 2.5 follows the motion in the video while rendering it in the look you want.

Specify which elements to reference.

Refer to the camera movement and motion in @Video 1...

Refer to the lighting changes, camera movement, and motion in @Video 1...

Write how the clay models correspond to the actual characters in the prompt as well.

Map the man in gray clothing from @Image 1 to the red model in @Video 1, and replace the green model 2 in @Video 1 with the red-haired girl from @Video 2.

A video connected to Seedance 2.5 is always used as a reference. Even if you use a word such as replace, the connected video is not edited directly; it is referenced to make a new video.

• Rough modeling is reflected better than detailed modeling.
• Keep elements such as trajectory lines and coordinate axes out of the clay model video.
• Even with a clay model video, describe the result you want in detail. Shortening the description because you have materials lowers the quality of the result.