SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsseedance-debugger

seedance-debugger

Testing & quality

View repository

Updated 10-15 17:32

About this skill

seedance-debugger

Input: the problematic prompt + the symptoms described by the user (or a video screenshot). Output: diagnosis (root cause) + fixed prompt + explanation of why to change it this way.

When not to use this skill

12-category problem diagnosis manual

See methodology/08-avoiding-12-pitfalls.md. A quick decision tree is given here.

Symptom identification → jump

Symptom stated by the user Jump
"The face doesn't look alike / face changes midway / looks like a celebrity" #1 character-id-drift
"Subtitles appear out of nowhere / there is text in the video" #2 subtitles
"There is a watermark / there is a Logo" #3 logo-watermark
"The anime style turned into a real person / the style changed" #4 style drift
"Frame jump at the seam / the transition is not smooth" #5 extension jump
"Two identical people appear" #6 twins
"After extension the quality becomes blurry / the face has mottled color patches" #7 quality degradation
"The effects are wrong / the countdown jumps randomly" #8 incorrect effects
"There are more/fewer characters" #9 too many reference characters
"There is a clicking sound at the end" #10 ending noise
"Chinese characters are read incorrectly" #11 Chinese pronunciation
"The timbre doesn't match the reference audio" #12 inaccurate timbre

1. Character ID drift

Root cause: mixed use of reference images (face + full body + clothing in one image) OR the face proportion is too small.

Fix strategy：

  1. Have the useradditionally prepare a headshot(containing only the face, preferably with no expression)
  2. Clearly distinguish in the prompt:<主体1> 的面部特征参考图片1（大头照），妆造参考图片2（全身照）
  3. Important materials first

Negative example：将图1中的人物作为主角... Positive example：将@图片1（大头照）的面部特征作为主角面部锚点，@图片2（全身照）作为整体妆造参考。主角...

2. Subtitles

Root cause: the reference image/video contains text; vertical screens are more likely to trigger it.

Fix strategy：

  1. Add constraints:保持无字幕，避免生成任何文字或字幕
  2. If the reference material contains text → first use Seedream editing to remove it
  3. If business allows,prioritize horizontal screensfor generation

3. Logo / watermark

Fix strategy: add at the end of the prompt 不要生成水印，不要生成Logo

4. Style drift

Root cause: the style of the reference image is inconsistent with the target style + the prompt does not emphasize the style.

Fix strategy：

  1. Clearly specify the style constraint in the prompt:2D 日漫风格 / 3D 国风漫画 / 写实电影感
  2. A more precise approach: first turn the reference image into the target style, then generate the video

5. Extended jump

Root cause: Alignment error between the model's continued frames and the original video frames.

Fix strategy(post-production editing fix):

  1. Import the video into Jianying
  2. Previous video segmentdelete the last 6 frames
  3. Next video segmentdelete the first frame
  4. Repeat for all splice points
  5. When continuingend at the moment of the transition cut, and the next segment starts with the new scene after the ending cut

6. Twins

Root cause: Character three-view used as reference OR subject definition unclear.

Fix strategy：

  1. Clarify character subject associations：张三（对应图片1）将绿色存折扔向站立的李四（对应图片2）
  2. Add global constraints： 
    视频全程禁止出现外形、着装、配饰完全一致的人物，
禁止生成同款分身、双胞胎效果，
同一画面中仅保留单个对应人物，不出现人物重复复刻。
  3. Optimize reference materials: independent single-person photos,don't use three-views
  4. Simplify the prompt: don't stuff the entire script in

7. Image quality degradation

Root cause: During extended continuation, image quality decays progressively, especially in facial areas.

Fix strategy：

  1. White model relay method: First use Seedance to convert the original video into a white model video, then use it as continuation input
    • Reference prompt:将视频转为白色 3D 模型，人物统一为纯白 3D 模型，无色彩、无纹理、无阴影，纯白背景，结构稳定、运动流畅
  2. Prefer high-definition images as references
  3. Control the number of continuations to avoid repeated stacking

8. Incorrect effects

Root cause: Describing effects in text leads to model misunderstanding.

Fix strategy: Usereference videos to define effects, so the model accurately understands the form and motion logic.

  • Example:数字"2999"出场方式参考视频1

9. Too many reference characters

Root cause: Output stability decreases with >4 people.

Fix strategy：Step-by-step generation

  1. Group the characters (e.g., 6 people into 2 groups of 3), and generate images separately
  2. Use the grouped images from the first step as references to generate the final video

10. Ending noise

Fix strategy: Use Jianyingvolume envelopeto do an audio fade-out

  1. Import the video into Jianying
  2. Select the audio track → Audio → Volume envelope
  3. Near the end of the video, drag the last volume keyframe down to 0 to form a downward slope

11. Chinese pronunciation

Fix strategy: Replace easily mispronounced characters withcommon homophones with the same pronunciation

  • Example:鳞龙山 → 吃龙山

12. Inaccurate timbre

Fix strategy：

  1. Add detailed timbre feature descriptions in the prompt:使用@音频1低厚温润带细碎颗粒感中年男声的音色说"……"
  2. Keep the dialogue style close to the tone and expression style of the reference audio

Workflow

Step 1: Symptom collection

If the user only says "the effect is wrong," ask 3 questions (mandatory):

  1. What is the symptom(what did you see that you shouldn't have / what did you not see that you should have)?
  2. What was the original prompt？
  3. What reference materials were used(image/video/audio)?

Step 2: Diagnosis → Output

Compare against the 12 categories above and give a diagnosis report:

## 诊断

**症状**：[用户描述]
**对应问题类别**：第 N 类「问题名」
**根因**：[2-3 句解释]

## 修复后的提示词

\`\`\`
[修复版提示词]
\`\`\`

## 改了什么

| 原版 | 修复版 | 原因 |
|---|---|---|
| [片段] | [片段] | [说明] |

## 还可能的根因

如果上面修复还不行，可能是 [其他可能性]，可以再问我。

Step 3: Escalation recommendations when necessary

If the symptom is something like "prolonged image quality degradation" or "inaccurate timbre"prompts cannot solve 100% of the problem, clearly tell the user:

⚠️ This is a known limitation of the current Seedance 2.0; prompts can only reduce the probability of occurrence and cannot 100% avoid it. Suggestion: [specific post-production remediation plan]

Practical examples

Example 1: Twins (the most common pitfall)

User input：

I had two characters in a scene together, but the generated image had four identical people. The prompt was this:

一个穿红色西装的男人和一个穿蓝色西装的男人在会议室争吵，桌上有文件散落，
他们都看起来很愤怒，镜头中景跟拍。参考图1是红西装男的三视图。

There is only one three-view reference image.

Diagnosis

Item Content
Symptom 4 identical characters appear in the image (there should be 2)
Corresponding category Category 6 "Twins"
Root cause (1) Three-view used as reference — the model treats different angles of the same person as multiple people. (2) Subject definition is unclear — the prompt does not have （对应图片X） clearly assigned. (3) Missing a global anti-duplication constraint.

Fixed version

将@图片1（三视图取正面）的男性定义为「张三」，穿红色西装；
另一名男性「李四」穿蓝色西装（无参考图，自由生成，与张三外貌差异显著）。

会议室场景。张三（红西装）与李四（蓝西装）隔着会议桌争吵，
桌面散落文件。张三右手拍桌，李四起身指向张三。中景跟拍，
镜头从张三背后绕到李四正面。

视频全程禁止出现外形、着装、配饰完全一致的人物，
禁止生成同款分身、双胞胎效果，同一画面中仅保留张三与李四两人，
不出现人物重复复刻。不要生成水印、Logo、字幕。

What was changed

Original version Fixed version Reason
参考图1是红西装男的三视图 @图片1（三视图取正面） Explicitly select a single viewpoint to prevent the model from interpreting the three-view as three people
一个穿红色...一个穿蓝色... 「张三」（对应图片1）...「李四」（无参考图） Clarify character-reference image attribution
(None) Added an anti-duplication constraint block Global constraint as a fallback
(None) Logo / 字幕 constraint Conveniently blocks the risks of categories 2 and 3

Other possible root causes: If duplication still occurs, replace "reference image 1" with a single-angle standalone photo (without using a three-view); the problem resolution rate is > 90%.

Example 2: Style drift

User input：

I wanted to generate a female protagonist walking in an anime style, but the output was a real person.

一个长发女孩走在樱花飘落的街道上，粉色头发，大眼睛，微笑。

The reference image is a real photo of this girl.

Diagnosis

Item Content
Symptom The target is 2D anime, but the output is a real person
Corresponding category Category 4 "Style drift"
Root cause (1) The reference image is a real-person photo, and the model follows the real-person style by default. (2) The prompt does not clearly constrain the style; it only describes the character's appearance features.

Fix strategy：Two-step approachMore reliable than only changing the prompt.

Step 1: First use Seedream to convert the real-person reference image into an anime-style image (example prompt:将这张照片转为 2D 日漫赛璐璐风格，保留人物五官特征，大眼睛、樱花色调）。

Step 2: Use the converted anime-style reference image together with the fixed prompt below:

2D 日漫赛璐璐风格，@图片1 的粉发女孩走在樱花飘落的街道上，
微风吹动发丝，樱花瓣缓慢飘落。中景跟拍，镜头从侧面缓慢推近，
最后停在她回眸微笑的特写。柔和粉色调，边缘清晰的线稿感，
扁平上色，无写实质感，无毛孔细节。不要生成水印、Logo、字幕。

What was changed

Original version Fixed version Reason
(Reference image: real-person photo) Convert to an anime image first, then use it as reference The style anchor must match the target style
(No style constraint) 2D 日漫赛璐璐风格...扁平上色，无写实质感 Explicitly lock the style, with both positive and negative constraints

Other possible root causes: If you only change the prompt without changing the reference image, the model still has a 50%+ chance of drifting back to a real person — the reference image is the main anchor for style, so you must change it first.

Example 3: Subtitles + watermark appearing together

User input：

A store sign subtitle inexplicably appears in the bottom right corner of the video, and there is also something like a watermark in the top left corner. Prompt:

雨天的咖啡馆窗外，一个女人撑伞走过，镜头透过雨水模糊的玻璃拍摄。

Diagnosis

Item Content
Symptom Bottom-right subtitle + top-left watermark
Corresponding category Category 2 "Subtitles" + Category 3 "Logo/Watermark"
Root cause The prompt has no constraint words at all. The "cafe" + "outside the window" context causes the model to automatically add store signs/slogans as environmental elements; vertical screen orientation is more likely to trigger subtitles.

Fixed version

雨天的咖啡馆窗外，一个女人撑伞缓步走过，镜头透过雨水模糊的玻璃拍摄。
雨滴沿玻璃斜流，女人身影在水雾中朦胧，背景街道灯光晕染成暖色光斑。
中景固定机位，景深浅，玻璃前景虚化。

保持无字幕，避免生成任何文字或字幕；
不要生成水印，不要生成 Logo；
玻璃窗及背景街道不出现任何店招、标语、品牌文字。

What was changed

Original version Fixed version Reason
(No constraint words) 保持无字幕，避免生成任何文字或字幕 Category 2 standard constraints
(No constraint words) 不要生成水印，不要生成 Logo Category 3 standard constraints
(No constraint words) 不出现任何店招、标语、品牌文字 Targeted patch for the "cafe" context

Other possible root causes: If the business allows it, switching to landscape orientation will further reduce the probability of subtitles (vertical 9:16 easily treats subtitles as poster layout).

Reference resources