NSFWJS is a client-side JavaScript library for detecting NSFW (Not Safe For Work) content in images using TensorFlow.js. It provides a simple API to load pre-trained models (MobileNetV2, MobileNetV2Mid, InceptionV3) and classify images into five categories: Drawing, Hentai, Neutral, Porn, and Sexy. The library is designed for both browser and Node.js environments, with backend selection guidance for optimal performance (WebGL, WebGPU, WASM, CPU). The core architecture consists of a lightweight `NSFWJS` class that handles model loading, inference, classification, and disposal, with support for tree-shaking via separate model imports. The repository also includes several demo applications: a full-featured React + Vite demo with Web Worker offloading, camera support, and file drop; a Node.js Express server example; and a minimal HTML/script tag demo. While primarily an application library, it is hybrid in nature as it bundles pre-trained model weights (quantized and base64-encoded) and includes model conversion scripts, making it both a software tool and a model distribution package. The library relies on TensorFlow.js as a peer dependency and uses IndexedDB for model caching in the browser.
This repository is a comprehensive frontend visual effects collection and demonstration platform. It contains over 100 self-contained demos and tools covering CSS animations, Canvas 2D/WebGL effects, Three.js 3D scenes, ECharts data visualizations, AI/ML integrations (TensorFlow.js for pose detection and body segmentation, MediaPipe, Tesseract.js OCR), and client-side media processing (image, video, PDF). The architecture is a React 18 single-page application (SPA) using TypeScript, Webpack 5, Ant Design, and react-intl for internationalization. All processing occurs entirely in the browser; there is no backend or persistent data store. The project serves as both a personal portfolio and a learning resource for advanced frontend techniques, but it is not a single-purpose production application. The code quality varies widely: some components demonstrate high craftsmanship (e.g., custom PNG encoder, image manipulation utilities, cross-page communication hook), while many others are boilerplate-heavy duplicates or exploratory demos. No automated tests are present, and the build output includes large minified chunks alongside source code, creating maintenance overhead.