SuperAPI AI Directory

seedance-troubleshoot

General & other

View repository

Updated 10-16 07:17

About this skill

seedance-troubleshoot · Diagnostic Tree (v5.0)

This skill diagnoses and fixes the most common Seedance 2.0 failure modes. It uses a simple diagnostic tree to identify the root cause and provides a specific, actionable solution.

The Diagnostic Tree

Start at the top and work your way down.

1. Is the output blurry, jittery, or morphing?

  • Root Cause: Overspecification. Your prompt is too long, has too many competing details, or too many actions in a short time (violating beat density).
  • Solution:
    1. Cut the prompt length to 30-100 words.
    2. Use a reference. Find an @Image or @Video that shows the style or action you want and replace 50+ words of description with a single @reference tag.
    3. Simplify the action. Reduce the number of distinct movements in the shot.

2. Is the camera chaotic, spinning, or ignoring your instructions?

  • Root Cause: You have violated the One-Move Rule by stacking multiple camera moves in a single shot.
  • Solution:
    1. Rewrite the camera instruction to include only ONE primary move (e.g., slow dolly push, handheld tracking, static wide shot).
    2. If you need multiple moves, use the [Cut to:] syntax to create a sequence of shots, each with its own single move.
    3. For complex moves, use a @Video reference and tell the model to Match the camera movement from @Video1.

3. Is the character not looking like your reference image?

  • Root Cause: The prompt is re-describing the character, which competes with and overrides the @Image reference.
  • Solution:
    1. Delete all physical descriptions of the character from the prompt (hair color, clothing, face).
    2. The prompt should only describe the character's action and emotion.
    3. Ensure the @Image reference is a clear, well-lit shot of the character.

4. Is the action stiff, slow, or lacking impact?

  • Root Cause: Lack of intent and physics language. The model doesn't know how to perform the action.
  • Solution:
    1. Add degree adverbs: violently, gracefully, frantically, dramatically.
    2. Add physics consequences: dust erupts, sparks fly, the character staggers, sweat flies off in slow motion.
    3. Use an @Video reference of a similar action to give the model a clear example.

5. Is the output just not what you wanted?

  • Root Cause: Your prompt is ambiguous. Words like cinematic, epic, and beautiful are subjective and mean nothing to the model.
  • Solution:
    1. Run the prompt through the Anti-Slop Check in the [skill:seedance-prompt] skill.
    2. Replace every subjective adjective with a measurable, observable detail.
    3. Iterate. Make small changes to your prompt and re-roll 3-5 times. Do not expect a perfect result on the first try.

CLI Debugging Tips

# Check task failure details
seedance status <TASK_ID> --json | jq '.error'

# List recent failed tasks
seedance list --status failed --json

# Retry with a different seed
seedance generate @prompt.txt --seed 2001 --wait

# Iterate rapidly: generate 5 variants, pick the best
for seed in 1000 1001 1002 1003 1004; do
  seedance generate "Your prompt here" --seed $seed -q
done

For full CLI reference, see [skill:seedance-cli].


Maintained by Emily (@iamemily2050)