A new AI model reads tomato leaves and actually answers questions about what's wrong with them, not just slaps on a label.
The model, called SOLAR, was trained to handle six different diagnostic tasks for tomato disease, from spotting symptoms to judging severity, using a single generative system. It combines a mixture-of-experts module to fuse image and text data, letting it reason across tasks instead of running separate classifiers for each one. The team evaluated SOLAR on 41,677 leaf images paired with 216,209 question-answer pairs, testing both closed-ended and open-ended queries. Across every task, SOLAR beat existing vision-only, vision-language, and task-specific models on accuracy and robustness.
Most plant-disease AI tools stop at a label: "early blight," done. SOLAR's pitch is that a farmer or agronomist can ask follow-up questions and get reasoning back, which matters more in a field than a bare classification. That's the harder, more useful problem precision agriculture has mostly sidestepped.
The code is on GitHub, which helps with reproducibility. Whether a mixture-of-experts visual-question-answering model holds up on a phone in an actual field, versus a curated 41,677-image benchmark, is a separate question nobody has answered yet.