Back to all articles

Seedance guides

Seedance Troubleshooting: Motion, Product Drift, Audio, and Failed Tasks

Diagnose weak motion, changing faces or packaging, missing audio, upload errors and queued or failed generations without blind retries.

By sdanceai product teamNov 11, 20264 min read

Start by separating a creative result from a task error. A finished video with weak motion needs a better brief or source image. A queued task needs monitoring. A failed task needs its status and error checked before a new submission. This guide gives a diagnosis order rather than promising that one prompt fixes every model.

The video barely moves

Check whether your prompt describes a visible subject action or only camera movement. “A beautiful rainy street” sets a mood; “the cyclist pedals from left to right while the wheels turn” defines motion. If the object is meant to stay still, request a camera push-in explicitly.

Try one action with one camera direction. Remove conflicting instructions such as “static camera” and “rapid orbit” in the same shot. A source frame with little room around a subject can make movement harder to preserve. Compare one revised attempt with the original instead of changing the model, source and prompt together.

Faces, labels or product shapes change

Use a sharp reference with complete edges and a readable subject. Remove abrupt turns, large camera arcs and scene changes. Ask for restrained motion and repeat the identity details that matter. An image cannot supply details hidden on the far side of a product.

Check the first, middle and last frames. For accurate packaging and lettering, preserve approved graphics or add them in your editor. Reusing references can improve guidance, but it does not guarantee identical faces, logos or proportions. For several shots, approve each clip before assembling the sequence.

The framing crops the product

Choose a compatible output ratio before submitting and leave space around the source subject. A vertical composition and a widescreen composition may need different reference crops. Do not ask a close-cropped source to reveal a full body without accepting that the missing information must be generated.

The result has no sound or the wrong sound

Check the selected model’s actual audio control, not just its family name. A prompt mentioning rain does not enable an unavailable or disabled audio option. If audio is supported and enabled, specify ambience, dialogue or silence clearly. Add approved music and captions during editing when precise delivery matters.

An image will not upload or a reference cannot be fetched

Use a raster format offered by the upload area and check its current size limit. A successful browser upload and a provider’s ability to fetch a reference are different checks. In an API workflow, localhost URLs, private intranet addresses, login-protected links and expired signed URLs are not reliable provider references.

Use a stable URL your generation service can access, with permission to process the asset. Do not publish confidential files merely to solve an access error. Reduce the number or duration of references only within the model’s advertised limits. See API input requirements for the distinction between web uploads and server-side media URLs.

A task stays queued or processing

Record the task ID and check its status. Queue pressure, model routing and the selected workload can affect waiting time; there is no fixed completion guarantee. Do not keep pressing Generate to repair a slow job. Each accepted submission can be separately billable.

If a submission request times out before returning an ID, its outcome may be uncertain. Check the workspace history or contact support with the approximate time and request context before creating another generation. The public API does not advertise an idempotency-key contract.

A task fails

The public task response may include taskInfo.userErrorCode. Useful categories include:

Code What to check next
reference_media_invalid Reference accessibility, format and size
reference_media_duration_invalid Per-file and combined reference duration
too_many_images, too_many_videos, too_many_audios Current model input limits
prompt_too_long Prompt length and repeated instructions
service_busy Wait and recheck before a deliberate new attempt
content_safety_rejected, copyright_restricted Review the content and rights; do not try to evade the restriction
generation_failed Keep the task ID and contact support if the cause is unclear

A failed or canceled task is not the same as a timeout in your own browser. Credit adjustments follow the task’s final state and the site terms; a local timeout does not itself prove a refund. Check the recorded balance before assuming credits have been returned.

Build a useful support report

Send the task ID, model, selected options, approximate submission time and a short symptom description to support@sdanceai.com. Include source media only when you have permission to share it. Do not send passwords, API keys or full authorization headers.

For a clean restart, follow the beginner guide. For better scene direction, use the prompt guide. For product work, use the advertising walkthrough.