🍃 Leaf disease detector

Upload a leaf photo. A TabPFN 3.5 tabular model reads 102 colour measurements and compares them with 7,500 labelled examples.

Tap to choose a photo — or drop one here

JPEG or PNG · one leaf, reasonably close up · max 12 MB your leaf

What this isno training

How it works, and how far to trust it

Method. Your photo is resized to 224×224 and reduced to 102 numbers — the mean and spread of colour in nine regions, two greenness indices, an 8-bin brightness histogram per channel, and left/right + top/bottom symmetry. TabPFN is a pretrained tabular model: it is not trained here, it just compares those 102 numbers against a fixed set of labelled examples drawn from the training split.

The friendly number. On held-out lab photos it scores —. Chance is 0.026.

The honest number. On — real field photos from a different lab, accuracy drops to —. Lab leaves are flat and evenly lit on a plain grey background; field photos are cluttered. Colour statistics partly measure the background, so treat this as a screening demo, not a diagnosis.

Deterministic. One estimator, fixed seed — the same image always gives the same answer, and your upload never changes the model.