It ranks what could be wrong, checks your local weather, reads extension guides and writes a care plan that cites them. When it isn't sure, it tells you to send a sample to a lab.
Free, no account. About 15 seconds for a quick look, a minute for a full investigation.
Models look, reason and write. Code makes the verdict and enforces the safety rules.
Plant, symptoms, where they sit, and whether the photo is good enough.
Up to three photos: a close-up, the whole plant, the underside. Kimi K2.6 reads them, because Token Factory has no NVIDIA vision model yet; Leaf switches when one appears.
Every plausible cause, each with a probability and the evidence for and against.
NVIDIA Nemotron 3 Ultra, with reasoning on. No cause is allowed above 90%, because a photo is never proof.
Looks closer, checks your weather, reads sources and can ask you up to three questions.
Nemotron Ultra calls tools: it crops a photo region and asks a targeted question, runs 14 days of your weather through published disease-risk rules (TOMCAST for tomato; Hutton, Smith and Wallin for potato), searches and reads extension pages with Tavily, and checks recent outbreak news. Code caps it at 8 tool calls, a dollar budget and 90 seconds. Tavily ranks .edu, .gov and extension sites first, and Leaf only reads pages its own searches returned.
A different model sees only the evidence and argues for the strongest alternative.
NVIDIA Nemotron 3 Super. It never sees Leaf's answer. If it disagrees strongly, the confidence drops.
Every step must cite a passage Leaf read. Steps the source doesn't support are removed.
NVIDIA Nemotron 3.5 Lightning writes it. A claim check then drops unsupported steps and any dose the source never states, and every product is checked against a table of 33 active ingredients for organic status, pets and bees.
Confident, Likely, or Send a sample to a lab.
From the numbers, not the model's wording: how far the leading cause is ahead, photo quality, the second opinion, and whether a serious disease is in play (backend/verdict.py).
Quick look first. Investigates only when it isn't confident, and tells you why.
15 s to 1 min
One careful look, a search and a cited plan.
About 15 s
The full investigation every time, streamed live.
About 40 s
Clear leader, good photo, nothing serious. A cited plan and a check-in date.
A favourite with a close second. The plan, what to watch for, and when to send a sample.
Too uncertain, or possibly serious. A printable case summary and your state's lab.
Ones where a wrong guess spreads past your garden: late blight, tomato brown rugose fruit virus, sudden oak death, citrus greening, Ralstonia race 3 biovar 2, boxwood blight, spotted lanternfly and box tree moth. They go to a lab when they're in the top two causes (at 15% or more) or reach 20% anywhere, second opinion included.
If your plan applies a product, Leaf checks the next five days of forecast hour by hour and picks the best 3 hours. One tap adds it to your calendar.
Fixed rules in code, Open-Meteo's hourly forecast. The product label always wins.
Tick "organic only" or "pets or children" and any step that breaks it is removed, with the reason. The model is told, and code checks anyway.
Each product is matched against 33 active ingredients and labelled organic or not, its risk to pets and whether it harms bees. Leaf also warns when two steps clash (oil and sulfur burn leaves together) and says how long to keep pets off.
Every plant you diagnose stays on your device for 30 days with its next check-in date. Add reminders to Google, Apple or Outlook calendars. No account.
My plants gives you a private link that copies the list to your phone. Anyone with that link can add check-ins, so send it only to yourself.
Send a new photo of the same leaves. Leaf says whether it's spreading, stable or recovering, and updates the advice.
When the answer is "send a sample", you get your state's plant diagnostic lab, how to collect and ship the sample, and a one-page case summary.
52 licensed photos: 31 diseases, 15 nutrient, water or sun problems, 6 pests.
14 of 16 times Auto said "Confident", it was right.
"Likely" is where the misses are, which is why it tells you what to watch for. Late blight is Leaf's most common miss.
| NVIDIA Nemotron 3 Ultra | First list of causes, and the investigator |
| NVIDIA Nemotron 3 Super | Second opinion, and fallback for every reasoning step |
| NVIDIA Nemotron 3.5 Lightning | Care plan, claim check, check-ins, chat |
| Kimi K2.6 | Reads photos, until an NVIDIA vision model is served |
| Nebius Token Factory | Runs every model call |
| Tavily | Search, page reading and outbreak news |
| Open-Meteo | Past weather for disease risk, forecast for spray windows |
| Cloudflare | Hosting, and saved reports in R2 |
No. It tells you when you need one, and hands you a case summary to take along.
Only small thumbnails are kept, for 30 days, under a private link. Your browser location is rounded to about 10 km before it's sent. Notes never appear on a shared link.
Common vegetables, fruit, roses and garden ornamentals. Weather-based disease rules cover tomato and potato. For anything unusual, expect "Likely" or a lab referral.
Yes. The project pays for the model and search calls, so the public demo has a daily limit. The code is MIT licensed.