Skip to content
← All articles

Counting calories with a photo: how accurate is it?

By Gautier, founder of Nutrinova · 17 November 2026 · 7 min read

Counting calories with a photo means letting an app recognise the food on your plate, estimate how much of it there is and convert that into energy and nutrients. The research so far says it works about as well as a person's own estimate: a 2023 systematic review of 52 studies found average calorie errors between 0.1% and 38%, depending on the method and the meal. This guide explains how the estimate is made, where it goes wrong, why even the label on a packet is an estimate, and what to look at when choosing an app.

How a photo becomes a number

Every photo-based estimate goes through three steps, whether the app does them one after the other or all at once.

  1. Recognition. The app identifies each food on the plate. In the 2023 review by Shonkoff and colleagues, 79% of the studies used a type of neural network trained on large collections of labelled food images. More recent apps also use general AI models that can read a whole scene.
  2. Portion size. This is the hard part. A photo is flat, so the model infers volume from cues such as the size of the plate, the cutlery or the depth of a bowl. In a 2025 study by Fridolfsson and colleagues, three general AI models given standardised photos with plates and cutlery as a reference all underestimated, and the error grew with the size of the portion.
  3. Conversion. The estimated weight of each food is matched to a nutrient database to give calories, protein, carbohydrates, fat and fibre.

Why oil and sauces are hard

A camera only sees surfaces. The oil in the pan, the butter melted into vegetables or the cream in a sauce leave little visual trace, yet they carry a lot of energy. According to the Swiss Food Composition Database, rapeseed and olive oil both contain 900 kcal per 100 g, so 10 g of oil adds about 90 kcal that may not show in the picture.

Studies point to this invisible fat as a weak spot, though not always in the same direction. In a 2026 Japanese study of 15 weighed hospital meals, all AI models overestimated fat by more than 20% on average. In a study of four popular apps by the US National Institutes of Health (NIH), presented at a conference in 2026, fat was underestimated by about 30 g per meal. The Shonkoff review also found smaller errors when a photo showed a single, simple food than a mixed dish.

What studies say about accuracy

The studies below compare AI estimates with a reference: food weighed in a kitchen or a lab, or nutrient tables.

StudyWhat was testedMain result
Shonkoff et al., 2023, systematic review52 studies of fully automated image methodsAverage calorie error 0.1% to 38%, similar to human estimates; not yet ready to stand alone in research or clinics
Fridolfsson et al., 20253 general AI models, 52 standardised photosAverage energy error about 36% for the two best; underestimation grows with portion size
Chen et al., 2025714 photos from 57 adults, against 3 days of weighed recordsEnergy mostly within 10 to 20%; fat and protein less consistent; adding the food's name helped
Isobe et al., 202615 hospital meals, AI against 10 dietitiansEnergy often within 10% for the best models and for dietitians; fat overestimated by over 20%
NIH conference abstract, 20264 popular apps, over 100 meals from a metabolic kitchenCalories about a third too low, 250 to 345 kcal per meal on average; preliminary, not yet peer-reviewed

Two patterns stand out. Energy is estimated better than individual macronutrients, and carbohydrates better than fat or protein. And results depend heavily on conditions: controlled photos taken from above do better than everyday snapshots with half the plate hidden.

Compared with your own estimate

Estimating by eye is not a neutral benchmark either. A 2022 systematic review by Whitton and colleagues, covering 29 studies and 2,964 adults, found that the most common errors in food diaries and 24-hour recalls were forgotten items and misjudged portions. Condiments were left out between 1% and 80% of the time depending on the study, vegetables between 2% and 85%. The Fridolfsson study concluded that the best AI models reached an accuracy comparable with traditional self-report, without the effort of writing everything down.

In short, a photo does not remove the uncertainty; it changes where it comes from.

Even the label is an estimate

The figure on a packet is not exact either. The European Commission's guidance on tolerances explains that declared values are averages, because foods vary naturally and change during production and storage. For a food containing 10 to 40 g of fat, protein or carbohydrate per 100 g, the measured amount may differ from the label by 20% either way; below 10 g, the margin is 2 g, or 1.5 g for fat. In Switzerland, the Federal Food Safety and Veterinary Office recommends following the same guidance, and treats products outside these tolerances as non-compliant.

So a weighed meal made from labelled products is closer to the truth than a photo, but it is still an estimate.

How to choose an app

No app is right for everyone, and none is exact. These questions help compare them.

  • Estimation method. Does the app work from a photo, a text description, a barcode or a food database, and can you combine them? Can you tell it about the oil, the sauce or the portion it cannot see?
  • Transparency. Does it say clearly that its numbers are estimates? Can you see what it recognised and correct a food or a portion?
  • Data protection. What you eat can reveal a lot about your health. Under Switzerland's revised Federal Act on Data Protection, in force since September 2023, data relating to health are sensitive personal data, and where consent is needed to process them, it must be explicit. Check where the data are stored, whether they are sold or used for advertising, and whether you can delete them. In the EU, the GDPR sets the framework.
  • Cost model. Some apps are free with ads, some keep the photo feature behind a subscription, some offer a trial followed by a subscription. Check what remains available after the trial and how to cancel.
  • Fit with your situation. For diabetes, kidney disease, a medical diet or a difficult relationship with food, a doctor or registered dietitian is a better starting point than any app.

Getting a better estimate from a photo

The studies above point to a few simple habits: photograph from above with the whole plate in view, keep foods visible rather than stacked, and add a few words for what the camera cannot see, such as "fried in a tablespoon of oil" or "with cream sauce". In the Chen study, giving the food's name alongside the photo improved accuracy.

Nutrinova works this way: you log a meal with a photo, a description or both, and it estimates calories, protein, carbohydrates, fat and fibre, with a Nova Score from A to E for each meal. The figures are estimates, useful for seeing the shape of a day rather than for counting to the gram. To put them in context, see how many calories you need per day and what a calorie deficit is.

Frequently asked questions

Are AI calorie counter apps accurate?
They give estimates, not measurements. In a 2023 systematic review of 52 studies, average calorie errors of automated image methods ranged from 0.1% to 38%, roughly in line with human estimates. Recent studies of general AI models and popular apps found that they tend to underestimate large portions and fat. Simple, visible foods are estimated best; oily dishes, sauces and mixed plates least well.
Are calorie labels accurate?
Not to the gram. Under the European Commission's guidance on tolerances, which Switzerland's food safety office recommends following, the declared value is an average, and a food with 10 to 40 g of fat per 100 g may contain 20% more or less fat than the label says. Natural variation, recipes and storage all play a part.
Which calorie counter app is best?
There is no single answer, as it depends on what you need. Useful questions: how does the app estimate a meal (photo, text, barcode, database), can you see and correct its estimate, where are your data stored and are they sold, and what does it cost once any trial ends. If you follow a medical diet, a dietitian is a better starting point than any app.
Is a photo better than estimating by eye?
Studies suggest the accuracy is similar overall. People mostly go wrong by forgetting items and misjudging portions, while AI mostly struggles with portion size and ingredients it cannot see. Adding a short description to the photo, such as the oil used, helps with both.

Sources

More articles

  • 17 November 2026 · 8 min read

    How many calories do I need a day?

    Your calorie needs depend on age, sex, size and activity. EFSA reference values in a table, the Mifflin-St Jeor formula, a worked example and how to check it.