Diagnosing Plant Problems
How to take great plant photos for AI diagnosis
A great plant photo can take a scan from "50% confidence, unclear" to "94% confidence, this is what it is and what to do." A bad one, too dark, too far, too cluttered, gives the model nothing to work with, and any answer it produces is a guess. Six rules below cover 90% of what makes a photo useful. They take about 30 seconds to apply.
1. Fill the frame with the affected leaves
The single biggest predictor of diagnosis quality is how much of the frame the plant occupies. If the diseased leaves are 10% of the photo and a couch is 60%, the model is working hard to extract information from a few hundred pixels.
Get the camera 6–12 inches from the leaves. The damaged area should fill at least half the frame. If you have a phone, tap to focus on the affected leaf so the camera doesn't focus on the background.
2. Shoot in natural daylight, no flash
Natural daylight gives the camera the best color information. Indoor lighting (especially warm bulbs or LED grow lights) shifts colors enough that yellowing, browning, and chlorosis become harder to read, sometimes for the model, always for you.
Best light: a few hours after sunrise or before sunset, near a north-facing window or in shaded outdoor light. Worst light: harsh overhead sun (causes strong shadows that hide texture) and any flash (washes out color).
3. Get both sides of the leaf when symptoms are subtle
Many problems are easier to identify from the leaf underside, spider mite webbing, scale insects, downy mildew, and the early stages of several rust diseases all live on the bottom of the leaf.
On Cultivator tier, GrowPilot supports up to 4 additional photos per scan. Submit one of the upper leaf, one of the underside, and one of the whole plant. The model fuses these into a single higher-confidence diagnosis than any one angle would produce.
4. Show the affected leaf in context, at least once
A close-up of a yellow leaf tells you what the leaf looks like, not where it is on the plant. If you have multi-photo support, add one wider shot showing whether the affected leaves are at the top, the middle, or the bottom of the plant. Location is half the diagnosis, bottom-leaf yellowing means very different things from top-leaf yellowing.
5. Keep the plant still
Slightest movement during a phone shot produces just enough motion blur to make fine textures (spider mite stippling, fungal threads, leaf miner trails) invisible. Hold the phone against a wall, rest your elbows on a table, or use both hands. Take 2–3 shots and pick the sharpest.
If you can see fine detail in the image when you pinch-zoom, the model can too. If you pinch-zoom and the image goes soft, retake it.
6. Crop out clutter before uploading
If the photo has a lot of background, counter, sink, other plants, crop it down to just the plant before uploading. The model spends some attention on every pixel; less clutter = more attention on what matters.
Most phone photo apps have a crop tool one tap away. Use it. A photo cropped to a tight rectangle around the affected area outperforms an uncropped shot of the same scene.
What about lighting setups, ring lights, macro lenses?
Nice if you have them, not required. A clean phone photo in good window light beats a poorly-composed shot under a $200 ring light every time. The model is trained on photos a typical grower would take, it's robust to ordinary phone hardware.
If you're doing this often (e.g., monitoring a production grow), a cheap clip-on macro lens for your phone ($10–15) makes a noticeable difference for diagnosing tiny pests like spider mites and thrips. Anything beyond that is overkill.
Frequently asked
- Does the model work on photos taken under grow lights?
- It works, but accuracy drops noticeably under strong purple or red LED grow lights, the color cast makes it harder to distinguish yellowing from healthy. If you can, briefly move the plant into natural light to take the diagnostic photo. If you can't, take the photo under the most neutral-colored grow light setting available, and the model will still produce a usable diagnosis with slightly lower confidence.
- What file format should I use?
- JPEG or PNG, anything under 10MB. Most modern phones default to HEIC, which GrowPilot also accepts, but if you have the option, JPEG is the broadest-compatibility format. Don't worry about resolution, modern phone cameras at default settings produce more than enough detail.
- Should I edit the photo before uploading (brightness, contrast)?
- No. The model is calibrated on unedited photos. Heavy edits, especially saturation and contrast boosts, can shift colors enough to flip a diagnosis between, say, nitrogen deficiency (uniform pale) and iron deficiency (interveinal). Crop if needed, but don't filter.
- How many photos do I need?
- One is enough for a basic diagnosis. For ambiguous cases, overlapping symptoms, multiple possible causes, Cultivator-tier multi-photo (up to 4 additional angles) consistently produces higher-confidence diagnoses. For everyday "what's wrong with this leaf" use, a single sharp close-up does the job.
- Can the model identify pests from photos?
- Yes, when the pests are visible, adult aphids, mealybugs, scale, caterpillars, leaf miners all show up clearly. Spider mites are harder because they're tiny; the model often spots them from the damage pattern (stippling, webbing) rather than the mites themselves. A macro lens or close phone crop helps.