What is wrong with this leaf?
Photograph one leaf. Get the likely crop and disease, the top 3 with confidence, and what to do next. The check runs on this device: free, no sign-in, and the photo is never uploaded. For a second opinion, an optional paid Expert check sends the leaf to an OpenAI vision model.
1. Your leaf
Drop a leaf photo here, paste one, or
One leaf, filling most of the frame, in daylight. Turn it over if the spots are underneath.
Drag the square onto the leaf. The model sees only what is inside it.
Picking the plant narrows the answer to that crop's conditions and is right more often.
See an example - instant, free
2. What it looks like
Your result appears here.
- Knows 26 diseases and the healthy leaf of 14 crops: apple, blueberry, cherry, citrus, corn, grape, peach, bell pepper, potato, raspberry, soybean, squash, strawberry, tomato.
- Best on one leaf against a plain background. Busy field photos are harder; check the top 3.
- It is a first opinion, not a diagnosis. Confirm with a local extension service or plant clinic before you spray.
Expert check - vision AI, uses credits
A second opinion from an OpenAI vision model: the crop, the likely problem (disease, pest, nutrient, weather or chemical damage, or healthy) and the signs it sees, 2-3 look-alikes and how to tell them apart, what to do, and when to call your extension service. It receives the square you framed (up to 768 px, no location data) and the top 3 above. Sign-in required.
Expert check 0 s
Your past expert checks - saved to your account
Recent checks - kept in this browser
Conditions it knows
How good is it?
On clean single-leaf photos like the training set it is right about 19 times in 20. On real field photos with soil, hands and other leaves in view it is weaker: the right answer is first about 2 times in 3 and in the top 3 about 9 times in 10. Picking the plant helps. When a photo looks unlike any training photo, it says so instead of guessing quietly.
How it works
A general image model (ViT-base/16, trained by Google on ImageNet) runs in your browser with onnxruntime-web, on the GPU when your browser offers WebGPU and on the CPU otherwise. The first run downloads it once (88 MB) and keeps it on this device. A small layer trained for this app reads the model's output and scores the 38 leaf conditions. It was trained on the PlantVillage photos, split so that no physical leaf appears in both training and test, plus the PlantDoc field photos. A second check measures how far the photo sits from every training condition; past the distance that covers 99% of test leaves, the result is marked unfamiliar.
The colour map is separate from the model: it sorts each pixel into green, yellowing, brown or pale by colour and reports the share of the leaf that is discoloured. Soil and shadows can count as brown, so look at the overlay.