返回资讯中心
外部精选
软件工程
#开发者实时流

Posterlet 开源 4507 条生成海报提示词数据集

来源标题:Show HN: Posters – open prompt dataset for generated poster design

Posterlet 在 GitHub 上公开了 4507 条海报提示词规格,每条都包含生成对应图片的完整结构化文本及成品示例。这些规格按布局、颜色、排版等分段标注,可自由复制、修改和使用,便于复现或接入自己的生成流程。数据集按子集分文件存放,每条为一个 JSON 对象。

Hacker News 最新allanren2026/11/4 · 13 天前9 分钟阅读

以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。

4507 poster prompt specs — the exact structured text behind each image, paired with the finished example. Free to copy, adapt and ship.

Every poster on posterlet.com is generated from a structured prompt spec rather than a one-line description. Posterlet is an SI (superintelligence) poster generator, and this repository publishes 4507 of the specs behind its output — the exact text that produced each image — so you can see how layout, color, typography and subject are constrained separately, and copy the patterns into your own pipeline.

All prompt text lives in data/, one JSON object per poster, split into one file per subcollection. The galleries below are the browsable index: every thumbnail links to the live, editable version.

{"slug": "...", "title": "...", "collection": "halloween", "subcollection": "sale", "ratio": "9:16", "width": 1536, "height": 2752, "style": "3d", "tags": ["..."], "prompt": "[Format] ... [Colors] ... [Layout] ...", "image_md": ".../md/x.webp", "image_lg": "...", "image_full": "...", "page": "https://posterlet.com/poster/..."}

Most specs are split into labelled sections, which is what makes the output reproducible:

The infographic collection is the exception: those specs are a single narrative paragraph, because a whole-page diagram is easier to specify as one description than as independent blocks.

Collection Subcollection Prompts

Halloween Posters

593

Sales & Discounts 109 (3p)

Events & Parties 358 (9p)

Greeting & Decor 126 (4p)

Black Friday & Cyber Monday

431

Christmas & Holiday

356

Sales & Discounts 92 (3p)

Events & Parties 73 (2p)

Greeting & Decor 191 (5p)

Infographics

168

Product Thumbnails

1847

Beauty & Skincare 395 (10p)

Food & Drink 490 (13p)

Apparel & Accessories 169 (5p)

Electronics 119 (3p)

Home & Kitchen Appliances 133 (4p)

Home & Living 49 (2p)

Health & Wellness 57 (2p)

Baby & Kids 73 (2p)

General Products 362 (10p)

Gig Covers

1112

Web & App Development 337 (9p)

Design & Branding 207 (6p)

Marketing & SEO 100 (3p)

Data & Analytics 70 (2p)

Video & Animation 46 (2p)

Audio & Podcast 93 (3p)

Writing & Business 69 (2p)

Photo & 3D 59 (2p)

General Services 131 (4p)

Pages hold at most 40 images, so a large subcollection continues as -2.md, -3.md and so on.

A poster's permanent URL is its posterlet.com page. The slug is the identifier, so the link stays valid however these files are re-split, re-ordered or re-paginated:

https://posterlet.com/poster/halloween-event-schedule-infographic-poster

Every thumbnail links there, and each record in data/*.jsonl exposes the same URL as page.

Copy a spec, swap the [Subject] and [Headline] blocks, keep the structure. The section headers are the part that carries the weight — models follow [Layout] instructions far more reliably when the frame, palette and subject are already pinned down by their own sections.

Text rendering (headlines, prices, tags) is the usual failure point: keep quoted copy short, explicitly name the type style, and state the position twice — once in [Layout], once in [Headline].

This repository is generated, never hand-edited. tools/sync.py pulls every public poster from the posterlet.com API and rebuilds the files; a scheduled GitHub Action re-runs it weekly.

Prompt text is licensed CC BY 4.0 — use it commercially, just credit posterlet.com. Example images are SI-generated and remain hosted on posterlet.com; they are shown here for reference and are not included in the license.

正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。

打开原文
去社区讨论 →