AI Image Recognition

A vision model runs inside your browser and names what is in your photo — nothing is uploaded

The model

A model is downloaded once and kept by the browser, so a second visit starts straight away; after that recognition needs no network at all. Only one model is held in memory at a time.

The picture

Drop a photo here or click to choose one — JPG, PNG, WebP or GIF
The picture is read straight from the disk and never leaves the browser.

Settings

Both settings only change what is shown, so they apply to the result already on the page without running the model again. The best guess is always kept.

About DINOv2 and teaching machines to see

DINOv2 is a family of vision models released by Meta AI in 2023. Its striking feature is how it learned: instead of being shown millions of pictures with a label attached by a person, it studied a very large collection of unlabelled images and worked out for itself which parts of a scene belong together. The result is a general-purpose sense of what things look like, which can then be pointed at a specific job with very little extra teaching.

That general sense is useful well beyond naming objects. The same kind of model is used to find similar photos in a large library, to sort a gallery, to spot defects on a production line, to follow objects between video frames, and to estimate how far away parts of a scene are. Naming what a picture shows, as this page does, is the simplest of these tasks and a good way to get a feel for what such a model does and does not understand.

A confident answer is not always a correct one. A classifier can only choose from the categories it was taught, so an unusual subject is matched to the nearest familiar thing, and a cluttered photo with several objects may be labelled by whichever one dominates the frame. Treat the percentages as a rough measure of how strongly the picture resembles a category, and try clear, well-lit photos with one main subject for the best results.