hongye007/open-pose-costume

Open Pose & Costume is a bilingual reference studio for character design, drawing practice, and AI-assisted creation. It gathers pose, costume, and character libraries, lets you pair references on a workbench, and sends that combination to an image-generation service you configure yourself. The catalog is the reliable part; generated images depend on the configured model and service.

What it does

The README documents a catalog of 288 pose sets, 1,439 pose views, 1,407 costumes, and 1,499 character references—4,345 reference images in total—as of 2026-10-30, including extended categories (project README). You can search in Chinese or English by name, keyword, ID, or multiple terms, filter and sort categories, switch pose views, view thumbnails and originals, and keep favorites and pairings in the current browser. Share links preserve the language, filters, and selected references (project README).

The pairing and generation workflow

Pick a pose and a costume, then add them to the pairing bench. Optionally choose a character from the library or upload one; character references are meant to preserve identity and style, while pose and costume references control action and outfit. The composed description updates with your choices and can be copied for use elsewhere (project README).

For built-in generation, open settings and enter a service URL, API key, model, size, and quality. The default service is https://atomflow.vip, the default model is gpt-image-2, and the endpoint must accept OpenAI-compatible multipart POST /v1/images/edits with pose and costume originals in image[] (plus an optional character image). The browser calls the service directly, so the service must allow CORS and the site must be served over HTTPS or local HTTP. Both pose and costume are required before generation starts; after a result appears, you can edit the description and regenerate, or compare results in side-by-side or overlay views (project README). Results depend on the model, description, and service configuration, and fees are set by that service.

Configuration, privacy, and data limits

Settings are stored in the current browser. The API key is session-only by default; the "remember key" option stores it locally and can be cleared. Keys are not placed in share links or the repository. Locally uploaded character images (PNG, JPG, WebP, up to 10 MB) stay in page memory, survive closing the bench but not a refresh or page close, and are sent to the configured service only when you click generate (project README).

The site loads images directly from the upstream source, so availability depends on that site. Character metadata comes from AI Portrait Gallery, and those images are not bundled in the repository. Syncing updates upstream metadata without executing upstream web code or downloading images (project README).

Running it and licensing

Local development requires Node.js 22.12+; the documented workflow exits if the fixed port 5199 is occupied, and you can choose another port. GitHub Pages deployment runs tests and builds on pushes to main. The built dist/ folder can be hosted statically, for example on Cloudflare Pages (project README).

Code is MIT-licensed, but images and upstream data keep their own rights (LICENSE). Pose and costume sources permit reference, editing, and commercial use in works, but prohibit redistributing, reposting, or reselling the image data itself—even edited versions. The character source is marked CC0 1.0 while also requesting no republication of its images, and all images are AI-generated fictional people; check the source notices before reusing assets (SOURCES.md, third-party notices). A separate Codex skill performs the same pose+costume+character generation workflow through natural language (companion skill).