ziguishian/MxPage

MxPage – AI‑powered desktop workbench for e‑commerce product images

What it is – An open‑source Electron‑based desktop application that lets merchants (or anyone creating product listings) generate main product pictures, multi‑page detail‑page graphics and social‑media visuals using large language and image models. The tool runs locally on Windows or macOS, talks to an OpenAI‑compatible API for text‑to‑image generation, and stores all data on the user’s machine.

Key capabilities

Area	What you can do
Product page creation	Upload a reference photo, fill in known product facts, pick a target platform (Taobao, Pinduoduo, Douyin, Xiaohongshu, etc.) and language, then the app automatically produces a hero image, a full‑length detail page (3‑10 hero images, 4‑20 detail screens) and a mobile‑preview mock‑up.
Content planning	The built‑in LLM planner extracts selling points, usage scenarios, evidence, and any missing facts, prompting the user to confirm unknown values before generation.
Multi‑platform styling	Platform‑specific prompt presets adjust wording and visual style for each Chinese e‑commerce channel.
Batch processing	Queue whole folders of products, run them sequentially, view per‑item failures and resume where you left off.
Image translation	Translate existing product pictures into other languages (single, batch or folder) using a fixed translation prompt – no product‑planning agent needed.
Editing & delivery	After generation you can edit each image, compare original vs. translated versions, keep a history of edits, and export a ZIP of all assets or a single long‑scroll detail image.
Monitoring	Real‑time view of the current workflow stage, per‑image latency, API cost estimates (shown in CNY) and a 12‑month usage heat‑map.
User centre	Avatar, recent‑usage stats, model configuration UI and an “About” page.

How it works (V4 workflow)

PREPARE → PLAN → WAITING_INPUT → ART_DIRECTION →
GENERATE → IMAGE_REVIEW → SET_REVIEW →
COMPLETED / PARTIAL / FAILED

• The LLM returns a JSON plan and a storyboard. The first hero image is generated, then the remaining images are fetched in parallel (max 2 concurrent calls). All layout is performed locally: text blocks are placed, line‑breaks preserved, tables kept intact, and the final long‑scroll image is assembled without further LLM calls.
• Check‑points are persisted, so you can stop the run, fix a problem, and resume without re‑paying for already‑generated assets.
• Errors (timeouts, bad responses) are not auto‑re‑tried – the UI asks you before spending more credits.

Getting started

1. Download the appropriate ZIP from the Releases page (Windows x64, macOS ARM64 or Intel). The package contains Electron and a tiny Node.js backend; no separate Node installation is required.
2. Run the executable (MxPage.exe or MxPage.app).
3. Open the API Configuration pane (bottom‑left avatar) and enter the URL and API key of an OpenAI‑compatible service that supports structured JSON responses and image generation (e.g., the official OpenAI API, Azure OpenAI, or a self‑hosted model gateway).
4. Click Upload to add a reference product photo, answer the prompted questions (size, material, etc.), choose the target platform and language, and hit Start.
5. Watch the progress bar, edit any generated image if needed, and finally export the assets.

Technical stack

• Frontend – React + Next.js (served locally) inside an Electron shell.
• Backend – Node.js 24, Prisma ORM with a SQLite database (prisma/dev.db).
• AI integration – lib/ai contains provider adapters; the default workflow expects a model named GPT‑6 Luna for text planning and GPT‑Image 2.5 Flare for image generation, but any model that returns the required JSON schema works.
• Image processing – sharp for resizing/composing the final long‑scroll pages.
• Testing – Jest unit tests, workflow‑focused tests (npm run test:detail), and smoke tests for the desktop build.

License & community

• Source code is released under the permissive MIT license. The generated images and any third‑party assets remain subject to the terms of the underlying model provider and any copyright holders.
• The project is maintained by 灵矩绘境 (MatrixInspire). Issues, feature requests and bug reports are welcomed via the GitHub Issues tracker; a QR‑code for a community chat is provided in the README.

Who should consider using it

• Small‑to‑medium e‑commerce sellers who need fast, consistent product visuals for multiple Chinese marketplaces.
• Marketing teams that want to prototype product pages without hiring a designer for every SKU.
• Developers interested in building or extending a locally‑run generative‑AI workflow (the codebase is fully open and includes the V4 state‑machine implementation).

---

All information above is taken directly from the repository’s README; no additional features have been inferred.