How Accurate Are Photo Calorie Counters? What the Research Says (2026)

In independent tests, photo calorie counters are usually off by roughly 20 to 35 percent on a single meal, and they tend to underestimate larger portions. That is better than most people’s unaided guesses and worse than a kitchen scale. The useful question is not “is it exact?” but “is the error small, random and correctable enough to track with?” This post walks through what the studies actually measured, what they did not, and how to use a photo tracker so the errors work in your favour.

By Naor Lugassi, founder of Kcaly AI. Last updated: October 29, 2026.

What counts as “accurate” here

Most studies report mean absolute percentage error (MAPE): the average size of the miss, ignoring whether it was high or low. A 30 percent MAPE on a 600 kcal meal means the typical estimate is about 180 kcal off. A separate question is bias: whether errors lean one way. For tracking, bias is worse than noise, because random misses cancel out over a week while a consistent underestimate does not.

Two caveats apply to everything below. Studies use different meal sets (restaurant plates, canteen dishes, standardized photos), and results on one cuisine do not transfer cleanly to another. And a number from a lab dataset is not the number you get from a blurry photo taken under a kitchen light.

The research, study by study

We are quoting figures from the papers as reported in secondary sources and abstracts; check the originals before relying on a specific number. Where we could not confirm a value, the table says so rather than guessing.

What the numbers add up to

1. Identification is mostly solved; portions are not

The 93 percent identification precision in the O’Hara study matches what we cover in AI food recognition accuracy: naming the foods on a plate is now the easy part. The error lives in estimating grams. A photo is a 2D view of a 3D object, and the same bowl of rice can be 150 or 300 grams.

2. Large meals are where the bias shows up

Both 2025 chatbot studies report errors that grow with portion size, in the direction of underestimation. That is the opposite of what most people want from a tracker during a cut: the biggest meals are also the ones logged most optimistically.

3. Humans are not a high bar

Unaided estimation is poor too. The Nutrition5k paper found raters and nutritionists missing by roughly 40 to 50 percent when judging calories from images, and older work on self-reported intake shows large underreporting (Lichtman et al., NEJM, 1992, found people underreported intake by about 47 percent on average in a specific population of self-described diet-resistant subjects). A photo tracker with a 25 to 35 percent miss is not perfect, but it is not worse than the alternative most people actually use. We make the same point in is AI calorie tracking accurate.

4. Noise averages out; bias does not

If your photo tracker is randomly high on Monday and low on Tuesday, your weekly total is much closer to the truth than any single meal. If it always misses the olive oil, your weekly total is wrong in the same direction every time. That is the case for the habits in the next section.

How to get a better number from the same photo

• Shoot from above, in daylight, with the whole plate and a fork or the plate edge in frame for scale.
• Say what the camera cannot see: “cooked in two tablespoons of oil” moves a stir-fry estimate more than a better photo does.
• Correct clear misses. If the photo says one cup of rice and you know it was two, fix it. Corrections are what remove systematic bias.
• Calibrate your eye occasionally. Weigh a few staples once; our guide to count calories without weighing covers when you can stop.
• Use a database for the lookup. Identifying the food with AI and pulling the nutrients from a verified source such as USDA FoodData Central removes one source of error that a pure language-model guess adds.

A first-party observation from WhatsApp

Because Kcaly AI runs inside WhatsApp, a correction is just a reply in the same thread: “that was half a portion” or “add a tablespoon of olive oil”. We see people correct more readily when it is a chat message than when it means opening a separate screen and editing a field. That is an observation from watching the product in use, not a measured result, and we have not run a controlled accuracy study. If you want to see how the flow works, start with how to track calories on WhatsApp and the AI food recognition page.

How this compares with other photo trackers

We have not run our own head-to-head, so we do not publish a ranking of competitors’ accuracy. For a feature-and-price comparison with one popular photo app, see Cal AI vs Kcaly AI photo calorie tracking compared, and for photographing technique, see log meals from photos. The sensible way to evaluate any tracker, ours included, is to log three or four meals you have weighed and see how far its estimates land from the scale.

Frequently asked questions

The bottom line

Photo calorie counters are accurate enough to build a habit on and not accurate enough to treat any one meal as exact. Expect a typical miss of roughly a quarter to a third, expect underestimates on big plates, and expect your weekly total to be closer than any single meal. You can try photo logging in WhatsApp and judge it against your own scale.