The rainbow slider used our photos. This one uses yours: Meta's DINOv2 ViT-S/14 — the actual network, 8-bit quantized to 23 MB — runs inside this tab on WebAssembly, and the famous PCA "rainbow" is computed in plain JavaScript. 🔒 nothing leaves your device
Wipe drags the divider · blend overlays the features · find similar: hover or touch a spot and every patch the model considers alike lights up.
Your photo was centre-cropped to a 518×518 square and cut into 1,369 patches of 14×14 pixels. DINOv2 turned each patch into a 384-number vector. We projected those vectors onto their 3 main directions (PCA) and painted them as red, green and blue. Similar colours = similar features.
The first principal component almost always separates "the thing" from "the background", so — like the DINOv2 paper — we keep only the foreground side and colour it with the next three components.
DINOv2 paper · code & weights (Apache 2.0) · DINOv2 on the map · runtime: onnxruntime-web (MIT).