How the Leafsight photo assessment works: what it shows and what it can't tell

A Leafsight assessment starts from one photo of the part of a plant that worries you. An AI vision model, Claude by Anthropic, looks at it and returns the likely problem, how serious it looks, how sure it is and what to try first. This page explains each step, what comes back, and what one photo cannot tell you.

How do you take the photo?

The answer can only be as good as the photo. The app's tips come down to three things:

Use daylight, avoid harsh glare and turn the flash off.
Fill the frame with the leaves that worry you, rather than the whole plant.
Keep it sharp: tap the leaf on screen to focus and hold the phone steady.

You see the photo before anything is sent. If the model finds it too blurry, too dark or with no plant in it, the app asks for another photo instead of guessing.

Before the check you can also answer a few optional questions: where the plant lives, how you water it, whether the pot drains, what changed recently and how long it has looked like this. Light, watering and drainage cause most plant problems, and a photo cannot show them, so the answers are weighed with the photo.

What does a result show?

Part of the result	What it tells you
The likely problem	A disease, pests or a care issue such as overwatering or low light, in plain words
How serious it looks	Healthy, mild, moderate or severe
How sure the model is	A percentage: the model's own estimate, not a measured accuracy rate
Visible symptoms and likely causes	What the model sees in the photo and what usually leads to it
Care steps	Treatment and prevention steps, non-chemical first, product types rather than brands, never a dose

When the model is less than 60% sure, the result says it is not sure and asks for one more photo, close up and in daylight. On a result with its care steps, you can ask follow-up questions, such as whether it can spread to your other plants.

If you save the plant, a later photo is compared with the earlier one, and the result says whether the plant looks better, about the same or worse.

What can one photo not tell you?

What is outside the frame: roots, soil, the pot, the room and the light it gets.
The history: how often it was watered, when it was repotted, what the weather did last week. The optional questions help with this, but only as far as you know the answers.
Problems that look alike. Several problems cause yellow leaves or brown spots, and the photo may not tell them apart.
Whether a plant or its fruit can be eaten or handled. Leafsight never answers that either way.

An assessment can be wrong, sometimes confidently, and no person reviews it before you see it. Treat it as the likely cause to check first, not a certain answer. For a valuable plant, or anything you might eat, ask a garden centre, a plant clinic or your local extension service.

Where does the photo go?

When you confirm, the photo is sent to Anthropic, whose Claude model makes the assessment. Leafsight's server checks the answer against safety rules before you see it. Photos are stored privately for your account, and you can download or delete your data in the app at any time; deleting your account removes your scans, photos and plants. The Privacy Policy has the details.

What is free, and what does Pro add?

You can take your first photo before making an account. A free account gets up to 5 assessments a day, each with the likely problem, how serious it looks, how sure the model is and the visible symptoms, and the first result that finds a problem includes its care steps. Leafsight Pro shows the treatment and prevention steps on every result and raises the limit to 100 a day.

Leafsight is free to download with free assessments every day, and with a free account the first result's care steps are free with a free account. After that, Leafsight Pro is $5.99 a week with a 3-day free trial, $4.99 a month, $29.99 a year or $49.99 once, or one result's steps for $2.99, one time. The price is on the screen before any trial starts. Every plan is on the pricing page.