How accurate is photo calorie counting? What the studies found

Photo calorie counting is fast, and it is often wrong by a lot. That is not a reason to give up on it. It is a reason to treat every photo estimate as a starting point you check, not an answer.

What independent tests found

A 2026 study by researchers at the US National Institutes of Health compared four popular apps with 102 meals prepared in a metabolic kitchen, where every ingredient is weighed. All four apps underestimated calories, by about a third on average. The average gap per meal ran from about 250 kcal for one app to about 345 kcal for another, and fat was underestimated by about 30 g per meal. The results were presented at the NUTRITION 2026 conference in July and are preliminary: they had not been through peer review when they were reported.

A 2024 University of Sydney study tested AI food tracking apps on dishes from several cuisines. Errors went both ways and were large on mixed dishes: the apps overestimated the energy in beef pho by 49 percent and underestimated pearl milk tea by up to 76 percent, as the university reported it. The apps did better with Western foods served as separate items than with mixed dishes, which are more common in Asian cooking, and the researchers asked app makers to train their models on more diverse and mixed dishes.

Peer-reviewed tests of the AI models themselves point the same way. A 2025 study in Current Developments in Nutrition gave three large AI models photos of weighed meals and found an average calorie error of about 36 percent for the best of them, with the models underestimating more as portions got bigger. One large lentil curry was estimated at 255 g when it weighed 480 g. A 2025 study in Nutrients found that adding an ingredient list to the photo cut the average error from about 31 percent to about 14 percent.

Why photos run low on home cooking

• Oil and ghee are mostly invisible. A tablespoon of oil is about 120 kcal and looks the same in a curry as a teaspoon. In the NIH study the apps missed about 30 g of fat a meal, which on its own is roughly 270 kcal.
• Big portions shrink in a photo. A deep katori and a shallow one look alike from above. The 2025 study found errors grew with portion size.
• Mixed dishes hide what is in them. Dal, stews, curries and soups can be thin or thick, with or without cream, and a photo cannot tell.
• Home recipes are not restaurant recipes. Most food databases are built around packaged and restaurant food.

How to log so the error stays small

1. Say how much oil went in. It is the single biggest correction most home cooks can make. See the hidden calories in ghee and oil.
2. Count what you can count. Two roti, three idlis, one paratha: counts beat guesses.
3. Fix the portion once. If your katori is bigger than the default, set it once and let the app remember. See katori, cup and plate sizes.
4. Save the dishes you cook every week as a recipe, so the photo only has to recognise the dish and not guess its contents.
5. Weigh now and then. A kitchen scale for a week teaches your eye more than any app.

What Homekcal does about it

Homekcal shows every photo estimate as a range with the likely value marked, not one number, until you confirm the food and the portion. Tap any number to see what it was based on: the matched dish, the portion, the oil level and the data source. Corrections take one tap and the next photo of that dish starts from your correction.

We have not yet measured Homekcal's own accuracy. We are building a reference set of weighed meals, starting from the public Nutrition5k set and meals people choose to weigh in the app. Until that is done and the method is published, we make no claim about how close our estimates are.