Beauty Generator
Overview
Use this skill when the user gives a short or fuzzy Chinese prompt for an adult female portrait and wants polished output without manually writing a long image prompt.
Keep the experience product-like:
- First give a compact core-choice round
- Then offer one optional detail round
- Send one short Chinese confirmation
- Generate two images directly
- After generation, offer an optional follow-up for a video-ready character reference sheet
Read these files only when needed:
Workflow
1. Scope and safety
- Only generate adult women.
- Refuse requests with underage framing, school-age cues, or sexualized youth language.
- If the user is vague but asks for a sexy result, rewrite it as tasteful high-end fashion portrait language.
- Keep the result realistic, elegant, and non-explicit.
2. Run the core-choice round first
If the user already gave some traits, keep them and ask only for the missing high-impact choices.
The first round covers:
- Preset template
- Face shape
- Eye shape
- Hairstyle
- Vibe
- Background
Use the short copy from references/conversation_flows.md. Encourage replies with letters, numbers, or short Chinese phrases.
3. Offer one optional detail round
Offer a second round for refinement only once. If the user skips it, use the defaults in references/option_catalog.md.
The detail round covers:
- Age feel
- Body tendency
- Makeup
- Hair color
- Expression
- Lighting mood
4. Resolve conflicts consistently
When user inputs conflict, use this order:
- Explicit user text
- User's later option choices
- Preset defaults
Do not silently replace a clear user requirement with a preset.
5. Assemble the portrait prompt
Build one polished Chinese natural-language prompt. Do not show the full prompt unless the user explicitly asks for it.
The portrait prompt should usually include:
- Adult identity and the chosen aesthetic direction
- Face shape, eye shape, hairstyle, hair color, and makeup
- Vibe, expression, age feel, and body tendency
- Background and lighting mood
- Realistic portrait-photography quality targets
- Clean facial detail, natural skin texture, no text, no watermark
Unless the user asks otherwise, keep the baseline:
- Default preset: 东方现代写实
- Half-body close portrait
- Clear sharp face
- Natural skin texture
- Clean high-end background
- Realistic photography
6. Enforce the portrait usage cap
Right before generating the two portrait images, run:
python .\scripts\usage_gate.py portraits consume
If the returned JSON contains allowed: false, do not generate. Return the exact message from the script.
If the user asks about remaining portrait uses, run:
python .\scripts\usage_gate.py portraits status
7. Generate exactly two portrait images
Use one coherent character design for both images and change only the angle:
- 正面微偏左
- 侧脸回望
Keep identity, styling, lighting direction, and overall aesthetic aligned across both outputs.
Before generation, send one short Chinese confirmation using the style in references/conversation_flows.md.
8. Offer the video-prep follow-up
Immediately after the two portraits are generated, ask a short follow-up in Chinese:
如果你要继续做视频素材，我可以基于第1张“正面微偏左”继续生成角色设定参考表。回复“需要”即可。
If the user does not say 需要, stop there.
9. Generate the character reference sheet on demand
If the user replies 需要, treat it as a request for a video-ready character reference sheet based on the first portrait image, not as direct video generation.
Before generating the reference sheet, run:
python .\scripts\usage_gate.py reference-sheet consume
If the returned JSON contains allowed: false, do not generate. Return the exact message from the script.
If the user asks about remaining reference-sheet uses, run:
python .\scripts\usage_gate.py reference-sheet status
Then generate one image using the first portrait image as the identity reference plus the fixed art-direction rules in references/reference_sheet_prompt.md.
The generated sheet must:
- Preserve the same face and identity as the first portrait
- Use a white pure background
- Present a professional high-end fashion character reference layout
- Show four aligned full-body views and multiple detail callouts
- Stay photorealistic and print-ready
Response style
- Keep the interaction in Chinese unless the user asks otherwise.
- Be concise, guided, and product-like.
- Do not dump long prompts by default.
- Confirm selections briefly, then generate directly.
作者微信：ddff9294
加好友备注来意。
有公众号，持续输出AI内容 ~
