GRAFT Model Answers Tomato Leaf Disease Questions From Photos

The short version

GRAFT is a new AI that answers plain questions about sick tomato leaves from photos. The team posted it on arXiv in October 2026. It uses a visual question and answer format instead of one fixed label. The paper is a preprint and is not yet peer reviewed.

What the researchers did

Khang Nguyen Quoc, Minh-Phuoc Tran, Gia-Han Truong, and Luyl-Da Quach published “A Multi-Modal Generative Model for Tomato Disease Leaves Understanding” as arXiv 2609.19555. GRAFT is short for a GeneRAtive multi-modal Foundation model for Tomato pathology.

The model uses the SCOLD vision model for leaf images. It pairs that with a LiquidAI LFM2.5 language model for answers. An Image Mixture of Experts block routes visual features into text. Residual fusion merges the streams.

Tasks include spotting symptoms, naming diseases, judging severity, and giving care style answers as generated text. The team hosts code on GitHub under EnalisUs GRAFT and shares weights on Hugging Face.

Photo: Karolina Grabowska / Pexels

What they found

Public README files stress visual question answering. One model can answer many prompt styles on the same leaf photo. The repo lists checkpoints and a test script for inference.

Exact accuracy numbers live in the arXiv PDF tables. The design goal is flexible dialogue for growers who need words about severity, not only a class code.

Mixture of experts routing means different image patches may activate different expert blocks. That design can save compute on simple leaves while spending more capacity on subtle lesions.

Why it matters for you

Tomato diseases spread fast in greenhouses and fields. Phone advisors that accept photos plus questions could reduce extension load if error rates are low enough.

Agtech apps that scout with drones or handheld cameras can compare GRAFT to simpler classifiers. Answer style models can explain why a leaf looks sick, which builds user trust.

Extension offices may pilot chat style tools if local experts review answers before farmers act. That human gate reduces harm from a wrong line of text.

Seed companies tracking trials should note which image sets trained the vision backbone. Domain shift remains the top field risk.

Photo: RODNAE Productions / Pexels

What to keep in mind

This is a preprint, not yet peer reviewed. Treatment hints are informational, not a substitute for local agronomist advice or pesticide law.

Models trained in one region may fail on new cultivars or lighting. Re test on your own images.

Do not automate high stakes spray decisions without human review.

Rain drops and soil dust on lenses can fool vision models. Field pilots should include messy real world photos, not only clean lab leaves.

Watch for field trials that report false alarm rates, not only accuracy on curated sets. Growers care about wasted sprays.

Coops should test GRAFT answers against local extension bulletins. Language models can sound confident while citing the wrong chemical name.

Drone vendors should test leaf blur from rotor wash. Blur can look like disease to vision models trained on sharp greenhouse shots.

Low bandwidth fields may need on device models. GRAFT size and latency numbers in the PDF matter as much as accuracy for rural deploys.

Seasonal workers may use low cost phones. Test the model on small screens with cracked lenses, not only flagship demo units.

Share false positive photos with agronomists each week. Their labels improve the next fine tune more than raw accuracy scores alone. Track season shift when plants move from vegetative to fruiting stages because color cues change fast.

Key takeaways

• GRAFT treats tomato leaf disease work as visual question answering with a generative stack.
• Combines SCOLD vision with LiquidAI LFM2.5 text and mixture of experts routing.
• Code and weights are public on GitHub and Hugging Face.
• Supports symptom, disease, severity, and care style prompts in one model.
• Verify metrics in arXiv 2609.19555 before commercial claims.

Sources

• Primary — arXiv, A Multi-Modal Generative Model for Tomato Disease Leaves Understanding (October 2026). Model and experiments.
• Secondary — GitHub, EnalisUs GRAFT (2026). Code, weights, and usage steps.