AI Food Scanner

What a photo can and cannot tell you

A meal photo carries two kinds of information reliably: what the food is, and roughly how much of it is on the plate. Those two are enough to estimate calories within a useful margin for most whole foods — grilled chicken, rice, fruit, eggs, vegetables — because identity and portion are both visible on the surface.

What a photo cannot see is everything absorbed into the dish. The oil a curry was finished with, the butter in a sauce, the sugar in a marinade and the density of a stew’s base are all invisible, and they are frequently the largest single term in the calorie count. This is not a limitation of any particular model. It is a limitation of photographs, and any tool that claims otherwise is overselling.

How to take a photo that scans well

• Shoot from directly above. A shallow angle hides depth, which is exactly the dimension portion size depends on.
• Get something of known size in frame — a fork, a standard plate, your hand. Scale is the single biggest input to the estimate and the easiest to supply.
• Separate the components where you can. Food that overlaps reads as less food than it is.
• Use even light. A hard shadow across a plate removes part of the meal from the picture.
• One plate per photo. A crowded table splits the estimate across dishes you may not be eating.

Where the estimate is weakest

Accuracy is not uniform across foods, and it is worth knowing which side of the line your meal falls on. In rough order from most to least reliable: whole single foods, then plated meals with separated components, then sandwiches and wraps, then mixed rice and noodle dishes, then curries, stews and soups, and last anything deep-fried or eaten in a restaurant where portions and oil are both outside your control.

When the model is unsure about an item it marks it, rather than returning a confident number it cannot support. That flag is the most useful thing on the results panel — a hedged estimate you can see is hedged is worth considerably more than a precise-looking one that is wrong.

A better way to use a number like this

Single-meal precision matters less than most people assume. If your estimate is 15% high on Monday and 15% low on Tuesday, the week is close to right, and the week is the unit that determines whether you gain or lose. What ruins a food log is not imprecision — it is the meals that never get logged at all, because logging was too much effort at the time.

That is the case for photographing a plate rather than weighing it: a rough number you actually record beats an exact one you skip. And the number that finally settles whether a deficit was real is not any of this — it is what your weight and your measurements do over three or four weeks.

Common questions

How accurate is calorie counting from a photo?

Treat it as a good estimate, not a measurement. A photo carries no information about what is underneath the surface, how much oil a dish was cooked in, or how dense a portion is — and published work on both people and software estimating calories from images finds errors commonly in the 20 to 30% range, sometimes more for mixed dishes. It is reliable enough to tell a 300 kcal breakfast from a 900 kcal one, and not reliable enough to settle a 50 kcal difference.

Is this food scanner free?

Yes, and it needs no account. There is a limit on how many scans one visitor can run in an hour, because each one costs us a model call — if you hit it, wait an hour or use the app, which is built for logging every meal rather than trying one.

What happens to my photo?

It is sent to the model that reads it and returned as text. We do not store the image, and it is not attached to any account, because there is no account involved. Nothing about the scan is cached — the response is explicitly marked not to be stored by any CDN between us and you.

Why did it get my meal wrong?

Usually one of three things. The photo was taken at a shallow angle, so portion size had nothing to scale against — shoot from above with a fork or hand in frame. The dish was mixed, and sauces, oils and fillings are the part a camera genuinely cannot see. Or the food is regional enough that a plain name fits it badly. When the model is unsure it says so next to the item, and that flag is worth believing.

Which foods does it handle worst?

Anything where the calories are hidden rather than visible: curries and stir-fries where the oil is absorbed, soups and stews with dense bases, composite baked goods, and anything eaten out where the kitchen was more generous than a home cook. Whole foods on a plate — meat, rice, vegetables, fruit, eggs — read far more reliably, because portion and identity are both on the surface.