Female Portrait Director Skill

Saw a tweet from 李岳 (@liyue_ai) that is mostly a number and a small surprise: his "female portrait director" skill passed 1,200 stars on GitHub, and the past half month added another 200-plus — with zero promotion in between.

He did not spin the number. The tweet just says it was a bit unexpected, that nobody had pushed it anywhere, and that the stars kept coming anyway.

The pitch in the same breath is plain: 20 different portrait styles, 6 image tools, easy to pick up, and a guided tutorial the first time you run it after installing.

What it actually does

The repository readme opens by ruling out the obvious guess: this is not a generic prompt collection. It locks explicit requirements, chooses one visual route, and directs the subject, action, clothing, scene, camera, lighting and finish as one photographable moment.

That directing verb is doing real work. The skill parses a request into face, body, outfit, scene, camera and pose, lighting, and filter modules — then expands the short parameters into a coherent scene instead of mechanically repeating them back. Output comes in a fixed four-part shape: locked parameters, module analysis, the final prompt, and negative constraints.

Routes live in eight folders — beauty, commercial, curve, fantasy, fashion, lifestyle, oriental, realism — and each request loads exactly one style file, so conflicting style keywords never collide.

Twenty styles, and where v1.6 came from

The full menu, from the first-use guide:

1. 清纯生活照  2. 纯欲曲线生活照  3. 都市时尚写真
4. 古风仙侠美人图  5. 电商服装模特图  6. 复古港风写真
7. 法式慵懒写真  8. 新中式东方写真  9. 活力运动写真
10. 旅行假日写真  11. 影楼精修写真  12. 东方丰腴写真
13. 清冷仙气古风增强版  14. 明媚华贵古风增强版
15. 超近景真实人脸人像  16. 古风贵女水光妆
17. 黑珍珠墨金CCD曲线生活照  18. 元气丰腴柔光CCD生活照
19. 冷白清透CCD曲线生活照  20. 低调电影感摄影

The six CCD and cinematic entries are the newest layer, and the quoted tweet tells how they got there: Codex had just reset its usage window without enforcing the five-hour cap, so he spent about an hour adding six new style routes and shipped the batch as v1.6 — cold-white clear CCD, soft-light CCD, black-pearl ink-gold CCD, dewy-makeup gufeng, ultra-realistic close-up, and low-key cinematic.

The six image tools

Around the style routes sits a tool registry for functional tasks, each a single file loaded only when its trigger words hit:

• prompt-optimize — refine an existing prompt without changing its locked parameters

• failure-diagnosis — why the output drifted, why every face looks the same, where the color shift came from

• parameter-recommend — suggested parameter sets for batch shooting across scenes

• safety-rewrite — compliance rewrite that keeps the legitimate aesthetic target

• image-to-prompt — reverse-engineer a reference image into observable visual features, guesses marked as guesses

• reference-image-generate — skip the prompt, keep an authorized face or product, generate directly

The last one is the interesting boundary: it must first classify what each reference actually is — person, product, style reference, or image to edit — before any generation happens.

Install, and an unusually honest compatibility table

One command: npx skills add pointing at the repo's skill folder, or a plain git clone into the Codex skills directory. First invocation with no parameters prints the V1.6 tutorial — all twenty styles, templates, combination rules, and a complete worked example.

The 62-file distribution package lists compatibility with five agents: Codex, Claude Code, Cursor, GitHub Copilot, Gemini CLI. The readme then does something rare — it footnotes its own table. "Installer-verified" means the CLI copied every referenced file into the target package. Only Codex was actually runtime-executed in the release environment; the other rows are packaging claims, not claims that every host provides an image model.

MIT licensed, with safety boundaries spelled out separately: fictional clearly-adult subjects by default, identity preservation only for the user or an authorized adult reference, and a flat refusal set covering minors, nudity, and deceptive identity content.

The replies

Two stand out. One user said she now uses this female-portrait skill to shoot men, and 李岳's reply was simply: any generated results? Show me.

The other is the best read on why the numbers moved without promotion: not promoting it and still gaining 200 stars means it hit a real need — almost nobody in the portrait space wants to hand-tune deployment, almost everybody wants one command that works. Once the skill format lowered the barrier, spread stopped depending on pushes.

Numbers

As of July 29, 2026:

• Main tweet: 171 likes, 28 retweets, 275 bookmarks, 25,100 views, 47 replies

• Quoted tweet (v1.6 release): 109 likes, 47,633 views

• Repository: 1.2k stars, 142 forks

Original tweet: https://x.com/liyue_ai/status/2082489124758720992 Quoted tweet: https://x.com/liyue_ai/status/2077401662449406349 Repo link (author's reply): https://x.com/liyue_ai/status/2077401782192681413 Repository: https://github.com/liyue-aigc/female-portrait-director