Seedance 2.0 prompt guide: give each reference a job.
Before writing the scene, decide which file supplies the subject, which supplies the setting, and which supplies the movement. Then describe one visible action. A short reference map makes it easier to spot conflicting instructions and revise the part that went wrong.
This guide focuses on Seedance 2.0 reference planning and the BytePlus ModelArk route. It is not a Seedance 2.5 parameter guide. Check the exact model and input mode in the service you use; a familiar product name does not establish the same controls across services.
1. Assign responsibilities before combining files
Imagine a small scene featuring a moss-green desk lamp. You want the lamp's appearance from one image, the desk layout from another, and a gentle camera move from a clip. “Use all these references” leaves the intended relationship unclear. Write down what each file contributes.
These are fictional assets, not uploaded or inspected files. Replace the labels with your own files. On a small screen, scroll the table sideways.
| Planned asset | Its job | A boundary to describe |
|---|---|---|
| Lamp appearance image | Green shade, stem, base, and material. | The surrounding room is not the requested setting. |
| Desk environment image | Desk layout, wall, and light direction. | Any other lamp in it is not the subject. |
| Camera reference clip | A gentle lateral camera move. | Its subject, room, and sound are not requested. |
Check that the references can work together. A close-up that hides the base cannot settle its shape. A room image lit from the right conflicts with a request for a matching left-side shadow. Choose the intended light and find a suitable reference before adding more instructions.
Keep this map beside the prompt. It is a planning aid, not a guarantee of exact reproduction. For a person across multiple shots, the character consistency guide adds checks for appearance, wardrobe, and continuity.
2. Match the names to the actual request
The official ModelArk Seedance 2.0 tutorial describes references by media type and number. Numbers start at one and follow the order of assets of the same type in the request's content array. An Asset ID is not a substitute for that name in the text.
For a hypothetical request ordered as lamp image, camera video, desk image, the lamp is Image 1, the clip is Video 1, and the desk is Image 2. The clip does not turn the desk into Image 3. If you reorder the images, check their names again before submitting.
Hypothetical request order: 1. Lamp appearance image -> Image 1 2. Camera movement clip -> Video 1 3. Desk environment image -> Image 2 Actual files and bindings: not supplied in this example.
A service may display selectable reference labels or @ mentions. Use its actual binding controls. Typing an invented filename, an unbound @ label, or the name of a missing file does not upload a reference. Confirm the final attachments and their labels in the destination service.
3. Write the shot around one visible change
These are original, fictional, unrendered examples. They have not been tested in a generation service. The first assumes that you have supplied and bound the three files in the hypothetical order above; do not submit it with missing references.
A. Use the clip for camera movement only
Create a new shot of the moss-green desk lamp shown in Image 1, placed on the desk in Image 2. Use Image 1 for the lamp's appearance and Image 2 for the setting and daylight. The lamp remains stationary. Use Video 1 only as a reference for the gentle sideways camera movement; keep the requested lamp and desk as the scene. End with the full lamp clearly visible. Use quiet room ambience.
The action here is the camera move. The lamp does not also rotate, light up, fold, and transform. Compare the result with each reference's assigned job: did the camera move as intended, and did another room or object leak into the scene? If the clip contains cuts or several moves, choose a simpler reference instead of asking for only its “best part.”
B. Write the movement directly
If a camera clip adds little useful information, remove it through the service's input controls. For this alternative, assume only the lamp and desk images are attached as Image 1 and Image 2.
Create a new shot of the moss-green desk lamp from Image 1 on the desk from Image 2. Use the first image for the lamp's appearance and the second for the setting and daylight. The camera moves slowly to the right while keeping the whole lamp in view. The lamp stays stationary, and the light remains steady. Quiet room ambience.
This alternative makes the camera intention readable without a third reference. It is not a claim that fewer inputs always produce better results. Choose inputs that communicate something the text alone leaves uncertain. The camera movement guide helps describe the direction and framing you want to evaluate.
4. Keep creative instructions separate from input roles
In ModelArk, ordinary reference images and explicit endpoint images use different roles. The current video task API reference documents reference_image, first_frame, and last_frame, and treats omni references and first-frame or first-and-last-frame inputs as separate, mutually exclusive scenarios. Writing “first frame” in prose does not set an API role.
If the assignment requires a chosen opening composition, decide whether you need an explicit first-frame mode before reusing the example. For a planned start and finish, use the endpoint preparation guide and check which controls your route actually exposes. Do not turn these creative examples into a request schema by guessing field values.
Also name the task. Creating a new shot while borrowing a camera move is a different request from changing an existing clip or extending its ending. For an edit, identify the base video and the change you want. For an extension, identify the base and the continuation. Verify support for that task in the selected version before sending it.
5. Review responsibilities, then revise one problem
- Confirm the attachments. Check actual files, per-type order, and destination labels.
- Check the subject. Compare the lamp's shade, stem, base, and material through the movement.
- Check the setting. Look for unwanted objects, changed light direction, or a room borrowed from the camera clip.
- Check the motion. Separate camera movement from accidental subject movement. Watch the whole take.
- Check the delivery. Inspect crop and sound in the exported file, then note one visible problem for the next revision.
Reference planning card — keep outside the generation prompt Service / exact model / task mode: Actual attachments and per-type order: Subject appearance source: Setting and lighting source: Motion source or written camera direction: One main action / ending composition: Duration, crop, and sound controls checked in service: Observed mismatch in the rendered take: One change for the next attempt:
If the camera behaves correctly but the desk changes, revise the setting assignment or its reference. Rewriting every sentence makes it harder to tell which change helped. When the work needs multiple shots, use the storyboard template to keep each shot's references and review notes together.
Use the Shotmelo prompt planner to edit the text, save a local draft, or export a brief. Its current preview uses local templates and does not analyze references or render images or videos. Bind the real files, generate, and inspect the take in your selected service.