🌿 LeafScan

Plant disease from a leaf photo — runs entirely in your browser, nothing is uploaded.

Click or drop a leaf photo here

Loading the model (≈ 32 MB, once)…

What this is. A ConvNeXt-T trained on lab photos (PlantVillage), then fine-tuned on web and field photos (PlantDoc), for 38 crop / disease classes of 14 crops. This page runs an int8-quantised copy of the model with ONNX Runtime Web.

How much to trust it. On held-out field photos it is right about 72% of the time. Below a confidence threshold fitted on field photos it says "Not sure": it answers about two thirds of field photos, and about 85% of those answers are right. It only knows these 38 classes and is not a substitute for an agronomist.

Beyond this page. The full project adds an assistant that answers your questions about the diagnosed disease from 28 Wikipedia articles, with a local language model that cites its sources, says "I don't know" when they do not answer, and flags sentences its sources do not support. Its models are about 10 GB, so it runs on your own computer (instructions in the repository).

Code, experiments and limitations: github.com/marconcini/leafscan · model: Marconcini/leafscan-convnext-tiny