skill-cover

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Skill-封面

统一管理多种封面风格。每种风格拥有独立的双比例基底和视觉合同；人物身份与动作作为共享资产复用。目前已注册

实战风格

。

Centrally manages multiple cover styles. Each style has independent dual-aspect-ratio bases and a visual contract; presenter identities and gestures are reused as shared assets.

实战风格

(Real-World Testing Style) is currently registered.

输入契约

Input Contract

必须取得主标题。接受以下可选参数：

```
style
```
：风格名称或别名；未指定时使用注册表的
```
default_style
```
。
```
subtitle
```
：副标题；缺省时保留副标题区域留白，不擅自编造结论。
```
tag
```
：左上角标签，默认
```
AI 实测
```
。
```
highlight
```
：需要强调的词组，可重复传入。
```
gesture
```
：动作预设，默认
```
zhengzuo
```
（正坐）。
```
reference_cover
```
：任意参考封面；提供时进入“参考封面换人换字”模式。
```
identity_image
```
：项目数字人/真人身份帧；参考封面模式必填。
```
avatar_video
```
：原始数字人 master；需要现场截取身份帧时使用。
```
output_dir
```
：用户指定路径优先。

默认同时输出 3:4 和 4:3。只有用户明确只要一种比例时才减少输出。

A main title is required. The following optional parameters are accepted:

```
style
```
: Style name or alias; when omitted, the registry's
```
default_style
```
is used.
```
subtitle
```
: Subtitle; when omitted, the subtitle area is left blank, and no conclusions are invented.
```
tag
```
: Top-left label, defaulting to
```
AI 实测
```
.
```
highlight
```
: Phrase to emphasize; this parameter may be passed repeatedly.
```
gesture
```
: Gesture preset, defaulting to
```
zhengzuo
```
(正坐).
```
reference_cover
```
: Any reference cover; when provided, the system enters “replace the person and text from the reference cover” mode.
```
identity_image
```
: Project digital-human/real-person identity frame; required in reference-cover mode.
```
avatar_video
```
: Original digital-human master; used when an identity frame must be captured live.
```
output_dir
```
: A user-specified path takes precedence.

By default, both 3:4 and 4:3 outputs are generated. Reduce the output only when the user explicitly requests a single aspect ratio.

风格与动作路由

Style and Gesture Routing

读取
```
references/style-registry.json
```
。
将用户说出的风格名称与
```
display_name
```
、slug 和
```
aliases
```
匹配。
用户未指定风格时使用
```
default_style
```
。如果参考图明显要求另一种已注册风格，则优先尊重参考图。
读取注册项的
```
reference_doc
```
，只加载所选风格的详细视觉合同。
读取
```
assets/shared/gestures/registry.json
```
和对应动作的
```
manifest.json
```
；风格只用
```
default_gesture
```
与
```
allowed_gestures
```
引用动作，不重复保存动作定义。
未注册的新风格或动作不得假装已存在；分别按“新增风格”和“新增动作”流程建立资产包。

Read
```
references/style-registry.json
```
.
Match the style name stated by the user against
```
display_name
```
, slug, and
```
aliases
```
.
If the user does not specify a style, use
```
default_style
```
. If the reference image clearly requires another registered style, prioritize the reference image.
Read the registry entry's
```
reference_doc
```
and load only the detailed visual contract for the selected style.
Read
```
assets/shared/gestures/registry.json
```
and the corresponding gesture's
```
manifest.json
```
; a style references gestures only through
```
default_gesture
```
and
```
allowed_gestures
```
and does not duplicate their definitions.
An unregistered new style or gesture must not be presented as already existing; create its asset package through the separate “new style” and “new gesture” workflows.

制作工作流

Production Workflow

判断使用“注册风格”还是“参考封面换人换字”；选择标题、副标题、标签和强调词。
使用内置
```
view_image
```
查看脚本返回的基底、身份图、动作图和风格参考图。

运行

scripts/build_prompt.py

生成每个比例的提示词与绝对参考图路径。例如：

bash

python3 scripts/build_prompt.py \
  --style shizhan \
  --title '史上最强国产开源模型，Kimi K3完全实测！' \
  --subtitle '真实项目全流程挑战｜综合能力超过 Opus' \
  --highlight 'Kimi K3' \
  --highlight '完全实测'

使用内置
```
imagegen
```
能力，每个比例单独调用一次
```
image_gen
```
。禁止将一种比例机械裁剪成另一种比例。
严格按脚本返回的顺序传入参考图：比例基底负责整体构图，共享身份图负责脸、发型和黑 T，共享动作图只负责姿态；动作 manifest 若提供比例构图参考，则下一张图负责人物位置、尺寸和露出范围；最后一张风格参考图负责视觉校准。动作 manifest 的精确构图规则优先于风格中的通用人物范围。
输出后逐张检查。文字错误时只做一次“仅修正文字”的定向编辑，并重新声明身份、动作、构图、背景和色彩不变。
用户指定目录时保存到该目录。属于现有视频项目时保存到项目的
```
封面/
```
，工程中间产物不得进入
```
成片/
```
。
运行
```
scripts/validate_cover.py
```
检查双比例与基础清晰度，再完成人工视觉质检。

Determine whether to use a registered style or reference-cover person-and-text replacement; select the title, subtitle, tag, and highlighted phrases.
Use the built-in
```
view_image
```
to inspect the base, identity image, gesture image, and style reference returned by the script.

Run

scripts/build_prompt.py

to generate the prompt and absolute reference-image paths for each aspect ratio. For example:

bash

python3 scripts/build_prompt.py \
  --style shizhan \
  --title '史上最强国产开源模型，Kimi K3完全实测！' \
  --subtitle '真实项目全流程挑战｜综合能力超过 Opus' \
  --highlight 'Kimi K3' \
  --highlight '完全实测'

Use the built-in
```
imagegen
```
capability and invoke
```
image_gen
```
separately for each aspect ratio. Mechanically cropping one aspect ratio into another is prohibited.
Pass the reference images strictly in the order returned by the script: the aspect-ratio base controls the overall composition; the shared identity image controls the face, hairstyle, and black T-shirt; the shared gesture image controls only the pose; if the gesture manifest provides an aspect-ratio-specific composition reference, the next image controls the presenter's position, size, and visible area; and the final style reference controls visual calibration. The exact composition rules in the gesture manifest take precedence over the style's general presenter-range rules.
Inspect every output. If any text is incorrect, perform one targeted “text-only correction” edit and explicitly restate that the identity, gesture, composition, background, and colors remain unchanged.
Save to the directory specified by the user. For an existing video project, save to the project's
```
封面/
```
directory; engineering intermediates must not be placed in
```
成片/
```
.
Run
```
scripts/validate_cover.py
```
to check both aspect ratios and basic clarity, then complete a manual visual quality inspection.

参考封面换人换字

Reference Cover Person-and-Text Replacement

当用户要求“截取我的数字人替换图中人物”“照这张封面改成当前项目”或同义动作时：

优先从项目的原始 HeyGen/avatar master 截帧，不从带圆形遮罩、字幕或其他 UI 的最终成片截取。找不到原始 master 时才使用最终成片，并记录降级。

运行候选帧脚本并用

view_image

查看 contact sheet：

bash

python3 scripts/extract_avatar_frames.py <avatar-master.mp4> \
  --out-dir <项目>/封面/avatar-reference

选择脸部清晰、头发完整、神态自然、无眨眼中间态的帧。身份帧只控制脸、发型、肤色和服装；参考封面只控制背景、材质、配色和编辑层级。

运行参考封面模式：

bash

python3 scripts/build_prompt.py \
  --reference-cover <参考封面.png> \
  --identity-image <选中的数字人帧.png> \
  --title $'深度测评\nQwen3.8 最新实测' \
  --subtitle '三个真实项目｜能玩、能交互、能验收' \
  --highlight 'Qwen3.8' \
  --highlight '最新实测'

按脚本返回顺序传图：Image 1 仅作风格/版式参考，Image 2 是唯一人物身份参考。不得继承 Image 1 的人物、眼镜、服装、标题或伪文字。
3:4 与 4:3 必须分别调用一次
```
image_gen
```
并独立重排；禁止把竖版机械裁成横版。
检查人物身份、全部文字、短边分辨率和精确比例；将最终图存入项目
```
封面/
```
，身份候选帧留在工程归档，不进入
```
成片/
```
。

When the user asks to “截取我的数字人替换图中人物” (“capture my digital human and replace the person in the image”), “照这张封面改成当前项目” (“adapt this cover to the current project”), or takes a synonymous action:

Prefer capturing a frame from the project's original HeyGen/avatar master rather than the finished video with a circular mask, subtitles, or other UI. Use the finished video only when the original master cannot be found, and record this degradation.

Run the candidate-frame script and use

view_image

to inspect the contact sheet:

bash

python3 scripts/extract_avatar_frames.py <avatar-master.mp4> \
  --out-dir <项目>/封面/avatar-reference

Select a frame with a clear face, complete hair, a natural expression, and no mid-blink state. The identity frame controls only the face, hairstyle, skin tone, and clothing; the reference cover controls only the background, materials, color palette, and editing hierarchy.

Run reference-cover mode:

bash

python3 scripts/build_prompt.py \
  --reference-cover <参考封面.png> \
  --identity-image <选中的数字人帧.png> \
  --title $'深度测评\nQwen3.8 最新实测' \
  --subtitle '三个真实项目｜能玩、能交互、能验收' \
  --highlight 'Qwen3.8' \
  --highlight '最新实测'

Pass the images in the order returned by the script: Image 1 is solely a style/layout reference, while Image 2 is the sole presenter identity reference. Do not inherit Image 1's person, glasses, clothing, title, or pseudo-text.
Invoke
```
image_gen
```
separately for 3:4 and 4:3 and rearrange each layout independently; mechanically cropping the vertical version into the horizontal version is prohibited.
Check presenter identity, all text, short-edge resolution, and exact aspect ratio; save the final images in the project's
```
封面/
```
directory. Keep identity candidate frames in the project archive and out of
```
成片/
```
.

共享资产

Shared Assets

黑 T 身份：

assets/shared/avatar/black-t-identity.png

动作总注册表：
```
assets/shared/gestures/registry.json
```
默认动作“正坐”资产包：
```
assets/shared/gestures/zhengzuo/
```
备选动作“摊手”资产包：
```
assets/shared/gestures/tanshou/
```
动作规范：
```
references/gesture-presets.md
```

身份图控制脸、发型、肤色和服装。动作资产包独立保存参考图、别名、用途、限制、比例兼容性与构图规则；动作图只能控制姿态，不得覆盖身份图的脸部控制权。比例构图参考只控制人物位置、尺寸和上身露出范围，不得复制其中的旧文字。

Black T-shirt identity:

assets/shared/avatar/black-t-identity.png

Gesture master registry:
```
assets/shared/gestures/registry.json
```
Default “正坐” gesture asset package:
```
assets/shared/gestures/zhengzuo/
```
Alternative “摊手” gesture asset package:
```
assets/shared/gestures/tanshou/
```
Gesture specifications:
```
references/gesture-presets.md
```

The identity image controls the face, hairstyle, skin tone, and clothing. Each gesture asset package independently stores reference images, aliases, intended uses, restrictions, aspect-ratio compatibility, and composition rules; a gesture image may control only the pose and must not override the identity image's control of the face. An aspect-ratio composition reference controls only the presenter's position, size, and upper-body visible range and must not copy any old text.

当前风格

Current Style

shizhan｜实战风格

shizhan｜实战风格 (Real-World Testing Style)

风格合同：
```
references/style-shizhan.md
```

3:4 基底：

assets/styles/shizhan/base/base-clean-3x4.png

4:3 基底：

assets/styles/shizhan/base/base-clean-4x3.png

3:4 参考：

assets/styles/shizhan/style-reference/style-reference-3x4.png

4:3 参考：

assets/styles/shizhan/style-reference/style-reference-4x3.png

Style contract:
```
references/style-shizhan.md
```

3:4 base:

assets/styles/shizhan/base/base-clean-3x4.png

4:3 base:

assets/styles/shizhan/base/base-clean-4x3.png

3:4 reference:

assets/styles/shizhan/style-reference/style-reference-3x4.png

4:3 reference:

assets/styles/shizhan/style-reference/style-reference-4x3.png

新增风格

Adding a Style

新增

<style-slug>

时完成全部步骤：

在
```
assets/styles/<style-slug>/base/
```
放入独立的 3:4 和 4:3 空白基底。

在

assets/styles/<style-slug>/style-reference/

放入两张已确认的风格成品。

新建
```
references/style-<style-slug>.md
```
，记录配色、构图、字体、人物位置和禁止项。
在
```
references/style-registry.json
```
注册显示名、别名、资产路径、比例构图、共享身份、
```
default_gesture
```
、
```
allowed_gestures
```
和核心提示语。

用

scripts/build_prompt.py --style <style-slug>

测试双比例提示词。

生成一次真实样例并通过比例、身份、文字、手部和视觉一致性质检。

不要复制一整套共享人物到每个风格目录。只有风格确实要求另一位人物时，才为该风格覆盖

identity

。

When adding

<style-slug>

, complete every step:

Place independent blank 3:4 and 4:3 bases in
```
assets/styles/<style-slug>/base/
```
.

Place two approved style samples in

assets/styles/<style-slug>/style-reference/

Create
```
references/style-<style-slug>.md
```
and record the color palette, composition, typography, presenter placement, and prohibited elements.
Register the display name, aliases, asset paths, aspect-ratio composition, shared identity,
```
default_gesture
```
,
```
allowed_gestures
```
, and core prompt wording in
```
references/style-registry.json
```
.

Test the dual-aspect-ratio prompts with

scripts/build_prompt.py --style <style-slug>

Generate one real sample and pass quality checks for aspect ratio, identity, text, hands, and visual consistency.

Do not copy an entire shared presenter set into every style directory. Override

identity

for a style only when that style genuinely requires a different presenter.

新增动作

Adding a Gesture

先生成或截取动作候选并交用户确认。确认后在

assets/shared/gestures/<slug>/

建立独立资产包，至少包含

reference.png

与

manifest.json

；有明显比例差异时增加

reference-3x4.png

和

reference-4x3.png

。随后登记到

assets/shared/gestures/registry.json

，再由具体风格通过

allowed_gestures

引用。未确认动作不得标记为

approved

，也不得注册为默认动作。

First generate or capture candidate gestures and ask the user to confirm them. After confirmation, create an independent asset package under

assets/shared/gestures/<slug>/

, containing at least

reference.png

and

manifest.json

; when there are significant aspect-ratio differences, also add

reference-3x4.png

and

reference-4x3.png

. Then register the gesture in

assets/shared/gestures/registry.json

and have individual styles reference it through

allowed_gestures

. An unconfirmed gesture must not be marked as

approved

or registered as the default gesture.

交付质检

Delivery Quality Checks

至少检查：

两张图分别为精确 3:4 和 4:3，且是独立排版。
所选风格与注册表、风格合同和参考图一致。
黑 T 人物身份一致，位置、比例、手部结构合理。
所有文字逐字正确，无旧标题、伪文字或随机字符。
3:4 建议 1440×1920，4:3 建议 1920×1440；短边不得低于 1000 像素。
无水印、平台 UI、Logo 或未经用户授权的额外文案。
参考封面模式必须确认 Image 1 的原人物、眼镜、服装和旧标题均未残留。

通用提示词规则见

references/prompt-template.md

。

At minimum, check the following:

The two images have exact 3:4 and 4:3 aspect ratios and use independent layouts.
The selected style is consistent with the registry, style contract, and reference images.
The black T-shirt presenter identity is consistent, with reasonable positioning, scale, and hand structure.
All text is character-for-character correct, with no old title, pseudo-text, or random characters.
3:4 should be 1440×1920 and 4:3 should be 1920×1440; the short edge must not be less than 1000 pixels.
There are no watermarks, platform UI, logos, or additional copy that the user has not authorized.
Reference-cover mode must confirm that Image 1's original person, glasses, clothing, and old title have not carried over.

General prompt rules are documented in

references/prompt-template.md