portrait-916
人像生图默认 9:16 竖幅规则。TRIGGER: 人像生图、portrait generation、人像、portrait、人脸、9:16、竖幅人像、韩系人像、网红人像
- License 许可
- MIT
- 所属仓库
- rexleimo/aios
- 最近更新
- 2026-11-07
- Stars
- ★ 62
- Forks
- 11
01 前言 Front Matter
- name
- portrait-916
- tags
- [ "portrait", "image", "generation", "9x16" ]
- scopes
- [ "global", "project" ]
- clients
- [ "codex", "claude", "gemini", "opencode", "hermes", "workbuddy", "pi", "zcode", "qoder" ]
- description
- 人像生图默认 9:16 竖幅规则。TRIGGER: 人像生图、portrait generation、人像、portrait、人脸、9:16、竖幅人像、韩系人像、网红人像
- repoTargets
- [ "codex", "claude", "gemini", "opencode", "hermes", "workbuddy", "pi", "zcode", "qoder" ]
- defaultInstall
- { "global": false, "project": false }
- installCatalogName
- portrait-916
02 资源文件 3
- 参考文档Workflows/Challenge.md8.1 KB
- 参考文档Workflows/Deconstruct.md5.2 KB
- 参考文档Workflows/Reconstruct.md8.6 KB
03 SKILL.md 正文
Portrait 9:16 — 人像生图画幅升格规则
生成人像（portrait）图片时，默认一律用 9:16 竖幅。这是实测结论，不是偏好。
核心规则
- 人像 → 9:16：
gpt-image-2用--size 1024x1792。不要用 3:4 直接出人像。 - 为什么：3:4 画幅下模型把人脸摊进更宽的画面，五官更扁、皮肤纹理更弱、"抓拍感"更差；9:16 对应手机竖屏自拍的真实取景，人脸占比和景深逻辑都更对。2026-09-12 kr-realism 任务实测：同风格下 3:4 出图（gen/probe 时期）人脸明显弱于 9:16 v4（见 rex-videos
.workspace/2026-09-12-kr-realism/对比记录）。 - 下游要 3:4 时：先 9:16 生成，再用 Python（Pillow）裁成 3:4，禁止直接 3:4 生成后凑合。
- 提示词里同步声明画幅：prompt 首句写
9:16 vertical ...，与--size 1024x1792保持一致。
尺寸对照（gpt-image-2 支持值）
| 画幅 | size | 用途 |
|---|---|---|
| 9:16 | 1024x1792 |
人像默认 |
| 1:1 | 1024x1024 |
头像/方形位、快速试 prompt |
| 16:9 | 1792x1024 |
横版场景/背景图 |
执行要点
- 执行器用
rexai-image-generation技能的 Python 版rexai-image.py（进度可见、超时 600s、熔断/内容策略自动重试）。 - 9:16 大尺寸出图常需 60-180s，属正常，勿中途 kill（job 在服务端继续跑，kill 会丢 job id 上下文）。
- 内容安全过滤是概率性的：同一 prompt 可能过可能拦；被拦时脚本自动重试，仍失败就替换高危措辞（实测 "one hand ... in foreground" 易拦，"leaning forward with hands on the bed edge" 稳定过）。
验证记录
- 2026-09-12：kr-realism 任务，ref-01 参考图拆解 → v4 prompt（9:16, 1024x1792）一次通过，8 项场景锚点全中，人脸立体度显著优于 3:4 对照图。
04 同仓库更多 Skills
versioning-by-impact
Use when completing a task and deciding whether repository changes require a semantic version bump and changelog entry before commit/push.
verification-loop
Evidence-before-assertions workflow. Use before claiming work is done, before release, and after any behavior change in scripts/skills/MCP.
typesafe-ai
Build AI-powered software with TypeSafe: small units of AI intelligence you can use like programming primitives. Its System One models, including Jev, turn natural language and application state into typed judgments and probabilities that code can combine. Use when a feature needs programmable common sense, when brainstorming what AI could make possible in an app, or when an LLM prompt-and-parse step could become a structured decision. Applications include routing, ranking, extraction, verification, and interactive experiences; these are starting points, not the limits. Read live docs and cookbooks to find useful patterns and discover new combinations.
skill-opt-lite
Train and optimize skill documents using the SkillOpt methodology — the agent acts as both target and optimizer, running tasks, reflecting on failures, proposing edits, and validating improvements in a self-contained training loop. TRIGGER: optimize skill, train skill, improve skill, skill training, skill optimization, SkillOpt, skill evolution, skill iteration loop, skill feedback loop, skill tuning.
skill-constraints
Enforce browser automation safety constraints — tool usage order, anti-detection rules, screenshot discipline, and human-in-the-loop checks. Use when setting up or verifying browser/MCP tool operations, NOT for every skill execution.
seo-geo-page-optimization
Use when improving a single page for SEO and AI answer-engine visibility (GEO), especially when you need faster first-screen conversion and clearer traffic funnels.