How accurate is AI calorie counting from a photo?

An honest answer, including the parts that are inconvenient for anyone selling one of these apps. This page explains what the estimate is made of, which errors are structural rather than fixable, and what the numbers are good enough for.

Updated August 18, 2026Free to use · no sign-up for the calculators

Short answer

Good enough to track a trend, not good enough to call a measurement. Identifying what is on the plate is largely solved; estimating how much is not, because a photograph records area rather than mass. Estimates are closest on simple separated foods and drift furthest on mixed dishes, sauces and fried food.

We do not publish an accuracy percentage, because we have not run a study we would be willing to have checked. A number without a stated protocol, sample and error metric is marketing, not evidence.

What the estimate is actually made of

A photo estimate is three separate guesses stacked on top of each other, and they do not fail equally.

  1. What the food is. Recognising rice, chicken, broccoli. This is the part modern models do well, and it is the part demos show.
  2. How much of it there is. The weak link, discussed below.
  3. What that food contains per 100 g. On the search path this is a database lookup and you can check it. On the photo path it is not: the model produces the figures itself — see the sources section.

Marketing tends to talk about the first step because it is the one that looks impressive. The error in your day comes almost entirely from the second.

Why portion size is the hard part

A camera records a flat projection. Mass has to be inferred from cues — the size of the plate, the cutlery beside it, how the food is piled — and every one of those cues can be wrong. A bowl of rice at 150 g and the same bowl at 300 g photograph almost identically from above. That is not a shortcoming of one product; it is what a single 2D image can support.

Two consequences worth knowing. First, a wrong portion is a proportional error: get the rice twice as heavy as it is and you are 200 kcal out from one item. Second, correcting the serving size afterwards is the single highest-value thing you can do, and it takes a few seconds.

Practical rule. Photograph from an angle rather than straight down, keep a fork or a hand in frame for scale, and correct the portion whenever the food is dense — rice, pasta, oil, nuts, cheese. Those five account for most of the drift.

Where the nutrition numbers come from

Here the two paths part company, and it matters. When you search for a food, the figures come from a public database and you can check them yourself. When you send a photograph or describe the meal in words, the model produces the per-100 g values directly — there is no lookup in either of those paths, which is one more reason a photo estimate is an estimate. The public sources behind search are:

SourceRun byCoversLicence
USDA FoodData CentralUS Department of AgricultureGeneric and laboratory-analysed foodsPublic domain
Open Food FactsNon-profit, contributor-maintainedPackaged and branded productsOpen Database Licence (ODbL)

The two behave differently. USDA entries are measured and consistent, but generic — "chicken breast, roasted" rather than a particular brand. Open Food Facts covers the packet in your hand, at the cost of being contributor-entered, so an occasional entry is wrong or incomplete. Where both exist, the laboratory-measured one is the safer bet. And where accuracy genuinely matters, search rather than photograph: that is the path with a traceable number behind it.

Why we do not publish an accuracy percentage

Every app in this category quotes one, and they are almost never accompanied by a method. A percentage is only meaningful with four things attached: how many meals were tested, how the true value was established, what counts as correct, and how the errors were distributed. Without those, "90% accurate" could describe anything from a careful study to twenty lunches in an office.

We have not run a study we would be willing to hand to someone else to check, so we do not quote a figure. When we do run one, the protocol and the full error distribution will be published here alongside the headline, not instead of it.

What to ask any app that quotes a number. How many meals? Compared against what — weighed portions, or another app? Within what margin does "correct" mean? And what did the worst cases look like? An answer that exists is worth more than a big number that does not.

What the numbers are good enough for

  • Good enough: spotting that lunch is twice the size you thought, seeing a weekly trend, finding the two or three meals that account for most of your intake, staying roughly on target without weighing everything.
  • Not good enough: clinical nutrition, managing a condition where exact carbohydrate counts matter, or any decision where being 200 kcal out changes what you should do.

The useful framing is consistency rather than precision. A log that reads 10% high every day still shows you exactly the right trend, and the trend is what you act on. A log with holes in it does not, however accurate the entries that survived.

How to make your own numbers better

  • Weigh your food for the first two weeks. Not forever — long enough to learn what 100 g of rice looks like on your own plate.
  • Correct the portion whenever the estimate is visibly off. Two seconds, and it removes the largest error term.
  • Log the same meal the same way each time, so a mistake at least stays constant.
  • Prefer typing it in for anything mixed — a stew, a curry, a smoothie. Those are the worst case for a photo and the best case for a search.
  • Judge yourself on weekly averages, never on a single day.

Please read. These estimates are not suitable for medical use. If you are counting carbohydrates to dose insulin, managing a condition where intake must be exact, or working with a clinician on a therapeutic diet, use weighed portions and the figures your care team gives you. Nothing here is medical advice; see our health disclaimer.

You can earn the full version instead of paying for it

If you want to see how close the estimates land on your own meals, that costs nothing to find out — the free plan has no card and no end date, and the answer will be more useful than any number on a marketing page.

  1. Create an account and log a few meals. New accounts open with a two-day Plus trial — three days if somebody invited you — and then settle onto the free plan: no card, no end date, one meal scan a day.
  2. Publish an honest post, story, write-up or short video about what you found — or invite three people who go on to actually use it.
  3. Send us the link. A published piece unlocks the higher tier; three qualified invites unlock 90 days of the full version.
Create a free account →

Your post has to say you got free access in exchange — that is an FTC requirement in the United States. Access is granted whether what you write is positive or critical. Full terms: promotional access.

Frequently asked questions

Accurately enough to be useful for tracking a trend, not accurately enough to be treated as a measurement. The identification of what is on the plate is the easy half. The hard half is portion size, because a photograph carries no depth and no weight — the same bowl of rice can be 150 g or 300 g and look almost identical. Expect a photo estimate to sit within a sensible range of the true figure for simple, separated foods, and to drift furthest on mixed dishes, sauces and anything fried.
A camera records area, not volume or mass. Depth has to be inferred from context — the plate, the cutlery, the framing — and any of those can mislead. This is why every honest photo-based estimate lets you correct the serving size afterwards, and why correcting it is the single thing that most improves your numbers.
It depends which path you used. Searched foods are looked up in two public sources: USDA FoodData Central, run by the US Department of Agriculture, public domain and covering generic foods measured in a laboratory; and Open Food Facts, a contributor-maintained database of packaged products under the Open Database Licence. Both can be checked directly. A photograph is different — there the model produces the per-100 g values itself, with no database entry behind them.
Typing it in is more accurate when you know what you ate and can weigh it. The photo is more accurate than the entry you would not have made at all, which in practice is most of them. The best use is a photo for speed, then a correction to the serving size where it matters.
Cooking fat you cannot see, sauces and dressings, mixed dishes where ingredients are hidden, and anything where the density is unusual — nuts, cheese, oils. A dry chicken breast next to visible vegetables is close to the best case; a stew is close to the worst.
Weigh food for the first two weeks so you learn what a portion actually looks like, correct the serving size when the estimate is obviously off, and log the same meal the same way each time. Consistency matters more than precision: a log that is consistently 10% high still shows you the right trend.