Portrait Prompt Atlas

GPT Image 2 人像处理提示词图鉴：集中收录、按场景整理，让需要的人快速找到合适的提示词。

中文说明 · English · 在线图鉴 · X：@dboy_yi2025 · 精选效果图 · Agent Skill · 语料清理与分类 · 对标说明 · 版权与署名规则

仓库内效果预览

提示词不只以 JSON 或长文本出现。下面直接展示四条精选记录的结果图；点击图片可进入对应详情页，查看原文摘要、分类、身份保持状态和 X 来源。完整的八条精选案例见效果图画廊。

老照片与破损修复
_{GI2_01583 · @iamrealsnow}

自拍、脸部与近景精修
_{GI2_04109 · @iamahmedfaraz66}

插画、素描与动漫转换
_{GI2_21341 · @de_mon010}

棚拍美妆与近景肖像
_{GI2_21354 · @AiwithLariab}

图片来自仓库内 assets/images/，每张图都与一条原始提示词记录配对；来源发布者与作者核验状态不会被省略。

微信公众号宣传物料

这是维护者的公众号宣传图，和案例效果图分开存放，方便线下传播与仓库访客识别项目入口。扫描图中的二维码，或在微信里搜索“谁是专家”。

中文

这不是一个把提示词堆在一起的仓库，而是一套面向真人照片的可追溯图鉴：帮助你从真实需求出发，找到修复、自然美化、身份保持、职业头像和纪念照片处理的参考方案。

每条提示词都按来源原样保存：中文仍是中文，英文仍是英文；本仓库不把它们翻译成另一种语言，也不把发布账号未经核验地当成原创作者。

在线浏览

访问 Portrait Prompt Atlas 在线图鉴，可按处理方向、具体场景和身份保持强度筛选，查看结果图、原始提示词与 X 来源链接，并一键复制。

当前收录

728 条原始 X 来源记录经范围筛选与近重复归并后，公开图鉴保留 519 条，日期范围为 2026-05-27 至 2026-08-12。
每条记录均有一张对应结果图、X 原帖链接、来源发布者和作者核验状态。
剔除 146 条不属于“真人参考图人像处理”的案例，并归并 56 个重复簇中的 63 条重复记录；原始池、自动判断、视觉巡检与归并审计均可复核。
8 个处理方向与 17 个当前有记录的具体场景，详见语料清理与分类。

致谢

特别感谢苍何创建的 awesome-gpt-image-2。它以开源图鉴、演示站和 Agent 使用方式呈现 GPT Image 2 提示词库，启发了本项目把已收集的人像处理资料整理为一个更专注于修复、自然精修与身份保持的开源库和演示站。

本项目独立收集、筛选和维护人像方向记录；不复制对方的提示词、文案、视觉设计或商业功能。具体边界见对标说明。

分类总览

方向	记录数	典型用途
修复与影像保全	4	老照片、清晰度、上色和破损修复
自然精修与美化	10	肤质、妆容、自然美化与轻修饰
职业头像与形象	5	职业头像、企业形象与商务照片
生活方式与旅行人像	15	旅行、街拍、日常与环境肖像
时尚与编辑肖像	153	时尚写真、杂志感与高定表达
身份锁定风格转换	173	插画、动漫、点云与艺术化转换
纪念、家庭与关系影像	4	家庭、伴侣、儿童与纪念照片
创意肖像与场景表达	155	电影感、人物海报、概念创意与场景重构

让 Claude Code 与 Codex 使用图鉴

仓库内置 gpt-image-2-portrait-library Skill。它不是单纯的安装说明：包内带有 519 条清理后的原语言记录、来源与身份保持字段、命令行检索、可复现的生成器和完整性校验。网站目录与 Agent Skill 由同一份研究主库生成，因此不会出现网站和 Agent 使用不同提示词版本的情况。

npx skills add mizzlelover/portrait-prompt-atlas --skill gpt-image-2-portrait-library --agent claude-code codex --global --yes --copy

安装后可直接提出：

使用 gpt-image-2-portrait-library，为一张旅行人像找一个自然、保留身份的处理方案。

Claude Code 也可使用插件市场：

/plugin marketplace add mizzlelover/portrait-prompt-atlas
/plugin install gpt-image-2-portrait-library@portrait-prompt-atlas

Skill 支持三种工作模式：从收录库中找原始提示词、在原文之外明确给出修改补丁、或在无合适来源时标记为全新创作。无论何种模式，均不翻译已有提示词，也不把未核验发布者写成原创作者。详见 Agent Skill 使用说明。

维护者更新研究主库后运行：

npm run curate:portrait-library
npm run generate:portrait-skill
npm test

GitHub Actions 会检查生成文件、提示词原文一致性、X 来源链接及 Skill 打包完整性。

署名与使用边界

状态	页面展示	提示词展示
`original_creator_self_claimed`	已核验原创作者、原始 X 帖文、证据	附署名公开
`repost_creator_credited`	原创作者优先、转载者其次、证据	附署名公开
`publisher_unverified`	来源发布者、原始 X 链接、未核验标识	附来源与更正渠道公开
`repost_creator_unknown` / `source_not_verifiable`	X 链接与来源未知标识	作为研究记录公开，不主张作者身份

创作者可随时提出更正或移除请求，详见署名与更正流程。

English

Portrait Prompt Atlas is a focused, searchable atlas for GPT Image 2 portrait editing, restoration, retouching, and identity-preserved transformations. It organizes original prompts by outcome and scenario so people looking for portrait-editing workflows can find the right case quickly. Prompts remain in their source language and link back to the X source, with a clear distinction between publisher and verified prompt author.

Browse the live catalog or install the included gpt-image-2-portrait-library skill for Claude Code and Codex:

npx skills add mizzlelover/portrait-prompt-atlas --skill gpt-image-2-portrait-library --agent claude-code codex --global --yes --copy

Research corpus

728 raw X-source records dated 2026-05-27 to 2026-08-12; 519 curated records are published after scope filtering, visual review, and duplicate clustering.
One matched result image has been downloaded for every record.
Every record has a publisher / authorship status and a rule controlling whether its full prompt may be published.
Current rights gate: no prompt is treated as publicly reusable merely because the X publisher is known.

The public corpus preserves the original X-source link, the source publisher where available, the attribution status, one source result image, and a correction path on every record. A publisher handle is labelled as a publisher, never silently promoted to “prompt author.”

Portrait-specialist categories

Category	Records	Typical use
Restoration & preservation	4	Old-photo, clarity, color and repair-oriented workflows
Natural retouch & beauty	10	Natural skin, beauty and light retouching
Professional headshot & brand	5	Headshots, corporate identity and business output
Lifestyle & travel portrait	15	Travel, lifestyle and environmental portraits
Editorial & fashion portrait	153	Fashion, magazine and editorial work
Identity-locked style transfer	173	Illustration, anime, point-cloud and art-direction transfers
Memory, family & relationship	4	Family, couples, children and memorial work
Creative portrait & scene	155	Cinematic, portrait-poster, concept and scene reconstruction

The category, image link, attribution state, risk flags, and rule-based evaluation are all queryable in data/prompts.jsonl and data/evaluations.jsonl.

What makes a record publishable

Status	Public display	Full prompt
`original_creator_self_claimed`	Original creator, original X post, evidence	Allowed with attribution
`repost_creator_credited`	Original creator first, reposter second, evidence	Allowed with attribution
`publisher_unverified`	Source publisher, original X link, explicit unverified label	Public with attribution and correction path
`repost_creator_unknown` / `source_not_verifiable`	Original X link and explicit provenance-unknown label	Public as a research record; never assert an author

Creators can request correction or removal at any time. See correction and removal.

Evaluation model

Each candidate is assessed through six distinct lenses, described in the evaluation model:

Identity preservation
Natural-retouch realism
Technical control and failure prevention
Task and commercial usefulness
Reference-image dependency
Rights and safety readiness

This prevents a visually impressive but identity-breaking style transfer from being ranked alongside a reliable natural headshot retouch.

Data layout

data/raw-prompts.jsonl is the immutable discovery pool; data/prompts.jsonl is the curated public corpus, with range and near-duplicate decisions in data/curation-audit.json.
data/image-manifest.jsonl maps every prompt record to one downloaded result image.
assets/images/ holds one result image per record for internal review.

The result images are retained together with their record-level source links so that a visitor can judge the prompt as a visual workflow rather than as isolated prose.

Acknowledgements

Special thanks to Canghe / freestylefly and awesome-gpt-image-2. Its open-source presentation of a GPT Image 2 library, live catalog, and agent workflow inspired the structure of this portrait-specialist library and its demo site. This project independently collects and curates its portrait records; it does not copy the reference project's prompts, copywriting, visual design, or commercial features.

The research source pool was discovered through Goku-OpenLab/gpt-image-2-prompts-datasets, licensed CC BY 4.0. That dataset-level license does not establish who authored each X prompt; this project therefore preserves individual X-source attribution and applies the stricter record-level publication gate above.