How to use Seedance 2.0: inputs, reference rules and a multi-shot prompt

Seedance 2.0 is a video model from ByteDance. This page follows the maker's own documentation on BytePlus ModelArk, checked on 2026-10-26 (the tutorial was last updated 2026-09-29), and adds prices from Runway's API page. Every rule below comes from a cited document.

How do you use Seedance 2.0 step by step?

Using Seedance 2.0 takes five steps: choose a platform, choose a model version, set resolution and length, add references, then write a prompt that names each file.

1. Choose a platform. Seedance 2.0 is sold by ByteDance's own Dreamina app and BytePlus API, and by resellers such as Higgsfield and Runway (compare where to use Seedance 2.0). Dreamina lists Seedance 2.0 and 2.0 Mini next to 2.5 in its model menu. The API route on BytePlus activates with a balance above USD 30, an AI Savings Plan at the USD 30 tier or above, or a Seedance 2.0 resource pack with available quota.
2. Choose a version. BytePlus lists three: Seedance 2.0 for the best quality, Fast for a balance of cost and speed, and Mini for the lowest cost. They support largely the same features.
3. Set resolution and length. Resolution values are 480p, 720p, 1080p and 4k, and 4K is limited to Seedance 2.0. Length is 4 to 15 seconds. Aspect ratios are 16:9, 4:3, 1:1, 3:4, 9:16, 21:9 and adaptive.
4. Add references. Add the images, videos and audio you want the model to copy. Rules are in the next section.
5. Write the prompt. Refer to each file by type and number, such as Image 1 or Video 1. Counting starts at 1 in the order the files were added.

Higgsfield's tutorial, dated 2026-06-24, suggests testing at 720p, changing one variable per try, and adding audio only on the final pass. Treat it as app-level workflow advice that works with any version.

What can Seedance 2.0 take as input?

Seedance 2.0 takes text plus 0 to 9 images, 0 to 3 videos and 0 to 3 audio files in one request. Audio must come with at least one image or video: text plus audio, and audio alone, are not accepted.

BytePlus calls this "omni reference". What the model takes from each type:

BytePlus also lists first-frame and first-and-last-frame image-to-video, video editing, and video extension (continuing a clip forward or backward, or joining up to 3 clips).

The BytePlus prompt guide says not to use the full limit. Too many assets make style conflicts and blurry subject choice more likely. Start with one or two references, as Higgsfield suggests, and add more only when the clip needs them.

How does Seedance 2.0 reference video work?

A Seedance 2.0 reference video tells the model what to copy about motion, camera and style, and you name it in the prompt as Video 1, Video 2 and so on.

BytePlus gives sentence patterns for this. For a reference: "Reference [the camera movement] in Video 1 to generate...". For editing, parts you do not mention stay unchanged by default. For extension: "Extend Video 1 backward, [new content], and then end with Video 1" keeps the original footage in the result. Without that wording, the output usually contains only the tail of the original.

Two limits from the same docs:

• Do not refer to a file by its asset ID in the prompt. Use the type-plus-number form.
• Extending a model-made clip repeatedly lowers image quality, especially on faces. BytePlus suggests limiting the number of extensions and starting from high-definition references.

What is the Seedance 2.0 multi-shot prompt format?

The Seedance 2.0 multi-shot format is a numbered storyboard where each shot says who, where, what they do and how the camera moves.

BytePlus's prompt guide says the model handles a complex video best as a timeline: Shot 1, Shot 2, Shot 3, in the order events happen. Its example has a man running through an alley, knocking over a fruit stand and climbing a wall. The guide also says that exact timing such as "0-3 seconds" is unstable, and that forcing durations may give odd results.

Template built from that guide (for more fill-in formats see the AI video prompt generator):

[Image 1] is the [role name]. [Image 2] is the [location].
Shot 1: [camera move, one only]. [Role name] [specific body action, slow and small]. [Where in the scene]. [Sound].
Shot 2: [camera move, one only]. [Role name] [next action, linked to the last]. [Sound].
Shot 3: [camera move, one only]. [Final action or freeze]. [Sound].
Look: [lighting, colour, style]. No subtitles, no logo, no watermark.

Rules the guide gives for filling it in:

• Name each subject the same way every time, and bind it to a file ("the tall man in Image 1").
• Use one camera movement per shot. Push, pull and pan together add instability.
• Prefer slow, small, connected movements over sprints, big jumps and rolls.
• Keep the prompt short. A full script pasted in as a prompt can confuse the model.
• Keep dialogue in one language. Avoid mixing Chinese and English except for names.

Can Seedance 2.0 use real faces?

Seedance 2.0 cannot take uploaded images or videos that contain real human faces, so a photo of a real person is blocked on the official route.

BytePlus states that the Seedance 2.0 series "do not support direct uploads of reference images or videos containing real human faces." It lists three ways around the block:

Do not upload a celebrity, a classmate or anyone who has not agreed. Copyrighted characters carry rights risks the platform rules do not remove. Dreamina, Higgsfield and Runway add their own content rules on top.

Why does a Seedance 2.0 clip drift or break?

BytePlus documents identity drift, duplicated characters, unwanted subtitles and quality loss on extension, and suggests a fix for each.

• Face does not match. Add a close-up headshot of the face alongside the full-body photo, and place the most important asset first in the prompt. BytePlus advises against multi-view character sheets, because the model may read them as several people.
• Twin characters. Define each character and bind it to its image. Add a closing line forbidding identical copies. This lowers the chance of twins.
• Subtitles appear. Landscape output shows subtitles less often than portrait. Generate wide and crop later if you can. Subtitles cannot be fully prevented, so crop or edit them out.
• Style flips to live action. When the reference is realistic and you want anime, state the style in the prompt.
• Jumps at extension joins. BytePlus suggests trimming 6 frames from the end of the first clip and 1 frame from the start of the next in an editor.
• Frame jumps in image-to-video. Crop the input image to the target ratio, or set the ratio to adaptive.
• More than 4 people. Stability drops. Make images of groups of up to 4, then use those as references.
• Click at the end of narration. Regenerate, or fade the audio out in an editor.

How much does a Seedance 2.0 clip cost?

On the Runway API, a 10-second Seedance 2.0 clip costs $1.40 at 480p, $3.00 at 720p, $4.00 at 1080p and $15.00 at 4K, as of 2026-10-26.

Runway lists per-second rates in credits, and a credit is $0.01. The sum is credits per second × seconds × $0.01.

The "3 tries" column prices one keeper in three generations, a sensible budget for a new prompt. Runway lists a minimum charge per generation (56 credits for Standard, 44 for Fast, 28 for Mini), which does not bite on a 10-second clip. Its Seedance 2.0 rows show no separate charge for reference files, unlike its Seedance 2.5 rows, which bill reference video by the second.

Other routes price differently. Higgsfield's June 2026 tutorial gives 36 credits, about $1.55, for an 8-second 720p clip on its app. Our Seedance pricing guide compares more platforms. The AI video cost calculator and the Higgsfield pricing guide cover more. For free options, see Is Seedance 2.0 free.

Should you use Seedance 2.0 or 2.5?

Seedance 2.5 allows clips of 4 to 30 seconds, while Seedance 2.0 allows 4 to 15, so use 2.0 for short clips and 2.5 for longer scenes.

Dreamina markets Seedance 2.5 as live worldwide and lists Seedance 2.0 and 2.0 Mini beside it. The steps on this page apply to 2.0. Version 2.5 has different limits. For example, BytePlus says 2.5 editing and extension tasks accept only the adaptive ratio. Check the 2.5 tutorial before you reuse these settings.