convez.ai
SkillMarketing & VendasAvançado
Jarad-z

Fashion Item Extractor

Jarad-z/fashion-item-extractor

Transforma fotos de look em catálogo de produtos com fundo branco

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill Claude que decompõe fotos de looks em itens individuais e renderiza cada um como imagem de produto em fundo branco (cropping + image-to-image via Gemini/Seedream). Automatiza o fluxo completo: detecção visual, geração de prompts otimizados e síntese de imagens de e-commerce.

Instalação

$npx convez-ai Jarad-z/fashion-item-extractor

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Gerar catálogos de e-commerce a partir de fotos de lookbook
  • Extrair itens de street style e runway para análise de moda
  • Criar assets de produto automaticamente de imagens de styling

Veredito

Ferramenta especializada e bem arquitetada para o nicho de e-commerce fashion. Integração clara com Gemini e Seedream, documentação técnica sólida (com referências de vocabulário e geração de imagem). Recomendado para equipes de produto/fotografia que automatizam pipelines de catálogo.

README

A tradução para o português deste README ainda está na fila.

Fashion Item Extractor

从一张穿搭 / 秀场 / 街拍图,逐件拆解出每一个单品,并一路做到电商白底单品图。 A Claude Skill that breaks any outfit photo down into its individual pieces and renders each one as a clean, model-free, white-background product shot.

License: MIT Claude Skill


这是什么

fashion-item-extractor 是一个 Claude Code / Claude Skill。 拿到一张有人穿搭的照片(秀场、街拍、明星图、lookbook 都行),它会像一个有经验的买手 / 造型师, 把这身造型里每一件可识别的单品逐一拆出来,再真正还原成可用的单品白底图。

它是 fashion-model-shot 的逆操作 —— 两者正好互为正逆,形成闭环:

穿搭整图 ──(fashion-item-extractor)──▶ 单品白底图
单品白底图 ──(fashion-model-shot)────▶ 模特上身图

默认三步工作流

拿到穿搭图,技能默认依次走完这三步(用户喊停才提前停下):

步骤做什么产物
① 拆解 + 写 prompt从上到下逐件识别,每件看准品类/版型/颜色/面料/细节/风格,并写一段白底单品 prompt单品清单 + 每件一段独立 prompt
② 截出实拍图scripts/crop_items.py 按看图估算的比例坐标把每件矩形裁出来,当视觉锚点每件一张原图截图
③ 生成白底图scripts/generate_whitebg.py 把「截图 + prompt」喂图生图模型,真正出成品图每件电商白底成品图

两个例外让流程提前停下:用户明说"只拆解 / 只要 prompt / 别出图",或出图所需的 API key/依赖没配(此时停在步骤二并提示如何继续)。

仓库结构

fashion-item-extractor/
├── SKILL.md                          # 技能主文件(Claude 读这个)
├── references/
│   ├── fashion-vocabulary.md         # 服装/布光/颜色/各模型差异参考词表
│   └── whitebg-image-gen.md          # 两家图生图模型的接口/参数/价格/坑
├── scripts/
│   ├── crop_items.py                 # 按比例坐标矩形裁剪每件单品
│   └── generate_whitebg.py           # 图生图:截图+prompt → 白底图(Gemini / Seedream)
└── evals/
    ├── evals.json                    # 端到端行为评测用例
    └── trigger_eval.json             # 触发词命中/不命中用例

安装

把整个 fashion-item-extractor/ 目录放进 Claude 的 skills 目录即可:

  • Claude Code:~/.claude/skills/fashion-item-extractor/(或项目级 .claude/skills/)
  • 之后在对话里说"帮我把这身穿搭拆出来"等,技能会自动触发。

脚本依赖按需安装:

pip install pillow                 # 步骤二:裁剪
pip install google-genai           # 步骤三:Gemini / nano-banana
pip install volcengine-python-sdk  # 步骤三:火山方舟 Seedream

出图模型与环境变量

步骤三(真正出白底图)需要其中一家的 API key,从环境变量读取,仓库内无任何硬编码密钥:

nano-banana (Gemini)Seedream 4.0(火山方舟)
环境变量GEMINI_API_KEYARK_API_KEY
单价~$0.039 / 张¥0.20 / 张
水印强制 SynthID(去不掉)可关
适用海外 / 英文 prompt国内默认(中文 + 便宜 + 可关水印)

接口细节、参数、价格与坑见 references/whitebg-image-gen.md。

单独跑脚本

技能会自动调用这两个脚本,你也可以手动跑。

裁剪(坐标用 0~1 比例写进 items.json,避免命令行内联 JSON 在 Windows 被吞双引号):

py scripts/crop_items.py <原图> <输出目录> items.json
[
  {"name": "01_outer_cardigan", "box": [0.18, 0.18, 0.82, 0.62]},
  {"name": "02_print_dress",    "box": [0.28, 0.22, 0.72, 0.60]},
  {"name": "03_white_shoes",    "box": [0.32, 0.86, 0.70, 1.00]}
]

出白底图(含中文 prompt 建议用 --prompt-file 从文件读,避开转义):

py scripts/generate_whitebg.py --provider seedream \
   --image crops/02_print_dress.png \
   --prompt-file prompts/02.txt \
   --out out/02_white.png

Windows 提示:脚本用 py 调用;含中文时先设 UTF-8(set PYTHONUTF8=1)防控制台编码崩溃。

Evals

evals/ 下两份用例可用于回归校验技能行为:

  • trigger_eval.json —— 触发词应/不应命中(含"拆代码""拆数据库记录"等反例,避免误触发)。
  • evals.json —— 端到端行为期望(含"用户喊停只要 prompt""没配 key 停在截图"等逃生口)。

注:evals.json 里引用的测试图(test-assets/…)未随仓库分发;自备穿搭图替换路径即可运行。

English Summary

A Claude Skill that turns one outfit photo into per-item, white-background e-commerce product shots through a 3-step pipeline:

  1. Decompose & prompt — identify every wearable item top-to-bottom and write one clean-product-shot prompt per item.
  2. Crop — crop_items.py rectangle-crops each item from the source as a visual anchor.
  3. Generate — generate_whitebg.py feeds crop + prompt to an image-to-image model (Gemini / nano-banana or Volcengine Seedream 4.0) to produce the final white-background shot.

API keys are read from GEMINI_API_KEY / ARK_API_KEY environment variables — no secrets are committed. It is the inverse of fashion-model-shot (single garment → on-model shot).

License

MIT © jarad-z

Ver o original no GitHub

Bom para

Relacionados