NSFW JS

NSFW JS is an open-source JavaScript library built to classify images as safe or not-safe-for-work without sending any data to a server. Classification runs entirely client-side, using a TensorFlow.js model that scores an image against categories and returns a probability-based verdict. Because the analysis happens in the browser, no image data leaves the user's device, which is a meaningful distinction for developers who need moderation without adding server-side processing or third-party data handling. The project is aimed squarely at developers rather than end users. It ships as a library to be integrated into a web app, forum, upload form, or any interface where user-generated images need a lightweight automated check before publishing. The homepage demo lets anyone drag and drop or tap to test an image, but the real product is the underlying model and API that gets embedded into another codebase. In practice, teams pick from multiple model sizes with different accuracy-to-file-size tradeoffs; smaller models load faster but classify less accurately, larger ones are more accurate at the cost of additional download weight. The project explicitly acknowledges it isn't perfect and is tuned to err toward flagging borderline images as unsafe rather than missing them, which reduces false negatives at the expense of more false positives. The maintainers frame this openly and invite the community to improve the model over time through two linked open-source projects: a data scraper that gathers examples of common misclassifications, and a trainer that produces updated models. NSFW JS sits in the content-moderation and content-detection space alongside server-side moderation APIs, but distinguishes itself by keeping everything local to the browser. It's a fit for developers who want a free, embeddable, privacy-conscious first-pass filter rather than a fully managed moderation service with dashboards, appeals workflows, or guaranteed accuracy SLAs. Teams with strict compliance needs or zero tolerance for false positives will likely need this as one layer in a broader moderation pipeline rather than a standalone solution.