Pagewright
LLM-orchestrated e-commerce detail-page (详情页) generator
用通用大模型,一键生成电商商品长图详情页
English · 中文
Demo · 演示
Selling a real product, end to end. Pagewright read the live North Face
1996 Retro Nuptse Jacket
page — which is Akamai bot-protected, so curl is blocked and it's fetched through a real
browser (tier-3) — pulled the real price, 700-fill down spec, six colorways, sizes and the
official benefit ratings + description, then rendered this bilingual 详情页.
Pagewright 读取了 North Face《1996 Retro Nuptse》线上商品页(站点有 Akamai 反爬,curl 被挡 →
经真实浏览器(tier-3)抓取),取到真实价格、700 蓬松鹅绒、6 配色、尺码与官方性能标签及文案,
渲染出这张中英对照详情页。
Real copy, specs & pricing shown above; the product photos are heavily mosaicked here to
avoid redistributing © The North Face imagery — at run time Pagewright fetches the real photos.
Reproduce from examples/north_face_nuptse (add photos per its
ASSETS.md). Other bundled example: examples/sa_infinity.
图中文案与数据均为真实信息;产品图已打码处理以规避版权(版权归 The North Face),实拍图在运行时抓取。
English
Give Pagewright a product URL or your own images + a description, and it produces a
polished, optionally bilingual long-image — the kind of detail page you upload to Taobao /
Tmall / Amazon / Shopify — rendered pixel-accurately from real assets, not AI-painted.
acquire ─▶ extract ─▶ enrich ─▶ compose ─▶ render ─▶ verify
(URL) (LLM) (optional) (HTML) (PNG) (LLM QA)
Why HTML rendering, not text-to-image
An e-commerce detail page lives or dies on data fidelity — the price, the spec numbers, the
brand logo and the size chart have to be exactly right. Today's text-to-image / 文生图 models
(Stable Diffusion, DALL·E, Midjourney, Flux, Nano-Banana…) are the wrong tool for that job:
- Text comes out garbled. Generated images mangle words and digits —
$159 turns into $IS9,
Chinese characters melt into glyph soup. You can't ship a price you can't trust.
- Logos & icons hallucinate. The model paints a plausible-but-wrong brand mark or tech icon.
*(This project's predecessor learned it the hard way: AI-drawn icons that didn't match the bra