A frozen retina vision-language model (FLAIR, 25 MB) runs entirely in your browser. Nothing is uploaded.
Preparing…
2Meet the data
The model was prompted with two classes, normal and diabetic retinopathy. The stream also contains images it should flag. Click any image to send it into the stream.
PROTON+ never sees these labels; colours are only for you. Each stream opens with the clinic’s own images, then shifted ones arrive.
3Stream & detect
Images arrive one at a time. PROTON+ builds a prototype bank from the stream itself and flags images that sit far from it, even when the softmax is confident.
upload unlocks when the model is ready
image appears here
Press play or click an image
Zero-shot label: —
fused score —bank 0 / 0warming up
Softmax confidence (MCM)—
How sure the model is about its label. High confidence looks in-distribution to MCM.
Prototype distance (PROTON+)—
1 − cosine similarity to the nearest class prototype in the bank.