About PicsTag

Why this exists

Image libraries fail for a dull reason: nobody fills in the metadata. Tagging is slow, repetitive work that pays off only after every image has been done, so it gets postponed forever and the library becomes a folder you scroll through.

AI is good at the boring half of that job — recognising what is in a picture and proposing words for it. It is not good at the half that matters, which is deciding whether those words are the right ones foryour catalogue. PicsTag splits the work along that line: the model proposes, you dispose.

Why it runs in your browser

Every comparable tool asks you to upload your images to a server. For unreleased product shots, client work under NDA, or anything covered by a data processing agreement, that upload is the whole problem.

So PicsTag does not have one. The models run in a Web Worker inside your tab, using WebGPU where it is available and WebAssembly where it is not. The site is a static page with no backend: there is no server that could receive your images, which is a stronger guarantee than any promise we could write here.

Under the hood

• Tags come from an image classification model (ViT), which returns its top predictions with a confidence score each.
• Captions come from an image-to-text model (ViT + GPT-2).
• Both run through transformers.js and are cached by your browser after the first download.
• An optional cloud mode can call a more capable vision model through OpenRouter, using an API key you provide and that never leaves your browser.

Contact

Feedback, bugs, or a feature you need: contact@picstag.app.