--- title: LeafScan emoji: 🌿 colorFrom: green colorTo: gray sdk: static app_file: index.html pinned: false license: mit short_description: Plant disease from a leaf photo, runs in your browser models: - Marconcini/leafscan-convnext-tiny --- # LeafScan Upload a leaf photo: top-3 diagnoses with calibrated probabilities, and a **"Not sure"** answer when confidence is below a threshold fitted on field photos. The model (ConvNeXt-T, int8 ONNX, 32 MB) runs **entirely in your browser** with ONNX Runtime Web: nothing is uploaded. About 72% right on held-out field photos; it answers about two thirds of them, ~85% of those right. A research demo, not a diagnostic tool. Code, experiments and limitations: [github.com/marconcini/leafscan](https://github.com/marconcini/leafscan) · Model: [Marconcini/leafscan-convnext-tiny](https://huggingface.co/Marconcini/leafscan-convnext-tiny) ## Beyond this page: a cited assistant (local only) The full project adds a question-answering assistant after the diagnosis. It retrieves passages from 28 Wikipedia articles, answers with a local LLM (Qwen3-4B) citing each source, says "I don't know" when the sources do not answer, and flags with ⚠️ the sentences an NLI model finds unsupported. Its models are ~10 GB, so it runs on your machine, not in this page: ```bash pip install -e ".[app,rag]" LEAFSCAN_ASSISTANT=1 python -m leafscan.app # photo → diagnosis → cited answers, Grad-CAM ``` What the evaluation showed: the assistant abstains on 86% of unanswerable test questions and 78% of its cited sentences are supported by their source; letting the diagnosis filter the search would lose 16 points of recall, because a wrong diagnosis removes the answer. Details in the [README](https://github.com/marconcini/leafscan#phase-2--rag-assistant) and the [research log](https://github.com/marconcini/leafscan/blob/main/docs/research-log.md).