How accurate are photo calorie counters?
Across 52 studies, systems that estimate calories from a food photograph averaged between 0.1% and 38.3% error against a weighed or table-calculated reference, and they did best on single foods and worst on mixed plates. A person is no better: in a test of 2,028 people estimating calories from photos, the mean absolute error was 57.9%, and five nutrition experts scored no better than the untrained participants.
Below are the numbers those two findings come from, what a photograph physically cannot tell a model, and which decisions this class of tool is actually good enough for.
What does "accurate" mean when an app advertises a number?
On its own, almost nothing, because there is no agreed reference. In the 2023 systematic review of this field, 51% of the studies built their reference from nutrient tables and only 27% weighed the food. An app that matches a database entry for a photographed banana can report a very high accuracy without ever having been tested against a scale.
Three questions turn a marketing number into a useful one:
- Accurate against what: weighed food, or a database entry?
- On which foods: single items, or plates with several things on them?
- Over which portions: standard servings, or the portions people really eat?
How accurate are photo-based calorie estimates?
The most complete answer available is a systematic review published in Annals of Medicine in 2023, which pooled 52 papers published between 2010 and 2023.
| What was measured | Result |
|---|---|
| Average relative error, calories, across 52 studies | 0.10% to 38.3% |
| Individual relative error, calories | 0.00% to 79.6% |
| Average relative error, food volume | 0.09% to 33% |
| Single food photo vs several foods in one photo | Errors and ranges smaller for single foods |
| Papers comparing AI directly against human assessors | 1 of 52, with AI "within the range of accuracy levels from human coders" |
Put that on a plate. A 30% error on a 700 kcal dinner is 210 kcal. Repeat it across three meals and you have wiped out a 500 kcal deficit without eating anything you did not log. That is the honest stake, and it is why the rest of this page is about which decisions survive that error and which do not.
Is a person better at this than a model?
No, and the study that shows it is the most useful thing to read before judging any app. A 2018 crowdsourcing study in the Interactive Journal of Medical Research asked 2,028 people to estimate the calories in 20 photographed foods of known content.
- Mean absolute error: 57.9%
- Estimates within 20% of the true value: 5.15 out of 20
- Five nutrition experts averaged 5 out of 20, statistically indistinguishable from the untrained participants, with an absolute error of 130.2 kcal or 55.3%
- Error by item ranged from 23.0% for a turkey sandwich to 241.0% for a green tea cake
- The bias was systematic: energy-dense foods were consistently overestimated, energy-sparse foods consistently underestimated
One more result from that study explains why this category exists at all. Averaging the guesses of 10 ordinary participants beat the best individual expert, a 36% relative improvement. A vision model trained on a large set of labelled meals is doing a mechanical version of the same thing.
Why does the same plate get two different answers?
Because a photograph does not contain the information the calculation needs. It has no mass. It cannot show the oil a vegetable absorbed in the pan, what sits under the top layer of a bowl, how much sugar is in the sauce, or whether the mince is 5% or 20% fat. The model infers portion size from the plate, the cutlery and the framing, so anything that changes energy density without changing appearance is invisible to it.
This is the mechanism behind the review's clearest finding. A single food photographed on its own gives the model a clean shape to reason about. A mixed, sauced or layered dish hides most of its calories, which is exactly where the reported errors climb. We saw the same thing when we tested our own model on weighed plates: every large calorie miss sat on a large miss in grams. It is also why Calso's home page says the figures are estimates rather than measurements, and why the app throws away an entry when the photo is not food.
Is typing food into a database more accurate?
The database part is accurate. The typing part is where the error moved. A 2020 validation study in the Journal of Medical Internet Research compared MyFitnessPal's calculations against a national reference database, with the food entries coded by dietitians:
- Energy: 1.3% overestimate, correlation r = 0.96
- Fat 1.7% under, carbohydrate 6.4% under, protein 7.8% under
- Sodium 51% under and cholesterol 77% under, with weak correlations for both
So for calories and macros, a well-coded database entry is reliable. Then a human uses it. In a 2013 BMJ study of 1,877 adults, 1,178 adolescents and 330 parents leaving fast food restaurants, two thirds underestimated the calories in the meal they had just bought, and about a quarter underestimated by at least 500 kcal. Mean underestimation was 175 kcal for adults and 259 kcal for adolescents.
Even the printed label has a legal tolerance. Under United States labelling rules, a food's calorie content is not a violation until it exceeds 120% of the declared value, which means a compliant 400 kcal ready meal can hold 480 kcal.
What error rate is good enough?
It depends entirely on the decision you are making with the number. This is the table we would want a new user to read first.
| Your question | Precision needed | Photo estimate |
|---|---|---|
| Am I eating more or less than last week? | Direction only | Good enough |
| Am I holding a 500 kcal daily deficit? | Roughly 10% per meal | Not on mixed meals |
| Did I hit 150 g of protein today? | Per-food weights | Weigh the protein, photograph the rest |
| Is this food safe with my condition? | Not a calorie question | Ask a professional, see our Terms |
The practical consequence is that consistency beats precision. A tool that reads 20% high every single day still tells you whether the week went up or down, because the same bias sits in both numbers. A tool you abandon on Thursday tells you nothing at all, however good it was on Monday. Pick the method you will still be using in a month, then remove its known blind spots by hand: add the cooking oil, and weigh the dense foods.
FAQ
Does a bigger meal make the estimate worse?
Usually, yes. The 2023 systematic review found errors were smaller for photographs of single foods than for photographs of several foods together, and the BMJ fast food study found that underestimation grew as the real calorie content of the meal grew. A plain grilled chicken breast is close to the best case. A curry with rice, sauce and a side is close to the worst.
Can I trust an app that advertises 95% accuracy?
Not without knowing what it was measured against. Half the studies in the 2023 review used nutrient tables as their reference rather than weighed food, and results were far better on single foods than on mixed plates. A high number quoted with no reference, no food list and no portion range is a marketing figure, not a measurement.
Should I weigh my food instead?
Weigh the calorie-dense foods, photograph the rest. Olive oil is 884 kcal per 100 g, so a 20 g misjudgment costs 177 kcal. Cooked broccoli is 35 kcal per 100 g, so the same misjudgment costs 7 kcal. Where you spend the effort matters more than whether you own a scale, which is the whole subject of counting calories without a food scale.
Is Calso more accurate than other photo apps?
We make no such claim. Calso reads one flat photograph with a vision model, exactly like the other apps in this category, and it inherits the same limits: it cannot weigh food, see under the surface of a dish or spot the oil in the pan. Our own FAQ says the estimates will sometimes be substantially wrong. If you want the comparison spelled out, see Cal AI alternatives.
Is a photo estimate good enough to lose weight?
For the trend, yes, provided you log every day. A tool with a steady 20% bias still tells you whether this week was higher or lower than last week, because the bias sits in both numbers. A tool you use three days out of seven tells you nothing, however accurate it is on those three days.
Why do two apps give different calories for the same photo?
Because neither one is measuring. Each infers a portion size from the plate and the cutlery, then matches the food to a different nutrition database. Two plausible portion guesses and two different database entries produce two different totals, and neither app has any way to check itself against the real weight.
Three things to take away
- Published average errors for photo-based calorie estimation run from under 1% to 38%, and the spread is driven by the food, not by the brand of app.
- Humans, including nutrition experts, are in the same error band or worse, so the honest comparison is not "estimate versus truth" but "estimate versus your own guess".
- Use the number for direction, not for arithmetic. Add the oil you cooked with, weigh the dense foods, and log every day rather than perfectly.
Calso says what it does not know
Photograph your plate, get an estimate of calories and macros against targets built from your body and your goal. The home page lists everything the app does today, and everything it does not.
Sources
- AI-based digital image dietary assessment methods compared to humans and ground truth: a systematic review, Annals of Medicine, 2023, 52 studies.
- Calorie Estimation From Pictures of Food: Crowdsourcing Study, Interactive Journal of Medical Research, 2018, 2,028 participants.
- Accuracy of Nutrient Calculations Using the Consumer-Focused Online App MyFitnessPal, Journal of Medical Internet Research, 2020.
- Consumers' estimation of calorie content at fast food restaurants, BMJ, 2013, 346:f2907.
- United States labelling tolerance for calories: 21 CFR 101.9(g)(5). Nutrient values from USDA FoodData Central.