How accurate are photo calorie apps?
Accuracy · Published October 2, 2026
They're good at recognizing what's on the plate. They struggle with how much is there, and especially with fat you can't see. Here's what the research found and what to do about it.
The NIH study
Researchers at the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) presented the findings at NUTRITION 2026, the American Society for Nutrition's annual meeting, on July 25, 2026.
| Meals tested | 102 standardized meals, plus more than 200 in a follow-up analysis |
| How they were made | In a metabolic kitchen, every ingredient weighed to 0.1 gram |
| Apps | MyFitnessPal, Lose It!, Cal AI and Appediet |
| Calories | 250 to 345 too low per meal, on average |
| Fat | About 30 g too low per meal |
| Carbohydrates | Estimated more consistently |
| Hardest diet | Low-carb ketogenic |
Follow the fat
Fat has 9 calories per gram. Thirty grams of unseen fat is about 270 calories: most of the gap. (That's our math, not the researchers'.)
A photo sees the chicken, the rice and the salad. It doesn't see the oil the cutlet soaked up in the pan, the butter in the mashed potatoes, the dressing pooled at the bottom of the bowl, or the fat marbled through the meat. That's why the high-fat keto meals were hardest, and why carbs, which are easy to see, held up better.
Recognizing food isn't the hard part
A 2024 review of 18 nutrition apps from the University of Sydney (Li et al., Nutrients) found that photo-recognition apps identified dishes well: MyFitnessPal got 97% right and Fastic 92%. But automatic energy estimates were inaccurate, and the authors called for better training on mixed dishes and culturally diverse foods.
That matters if you eat enchiladas, arroz con pollo or a plate of pernil, rice and beans. Layered, sauced, mixed dishes are exactly where a photo guesses worst.
Why a photo can't weigh your food
- It's flat. It can't tell how tall the pile of rice is or how deep the bowl goes.
- It can't see inside. Burritos, tamales and empanadas hide their fillings.
- It wasn't in your kitchen. It doesn't know whether the potatoes were baked or fried.
- Every recipe is different. Your mom's version isn't the restaurant's.
What to do with this
- Treat the photo as a first draft. Fix the portion if you know it.
- Add what the camera missed. A tablespoon of oil is about 14 g of fat. If you cooked with it, log it.
- Look at the week, not the meal. Trends are more reliable than any single number.
- Be suspicious of perfect precision. An app that says "523 calories" with no range isn't more accurate. It just isn't telling you.
How Micela handles it
Every dish on Micela is built from a reference recipe, ingredient by ingredient, with the grams in plain sight and the source of every number, so you can see where the calories in a burrito or enchiladas come from and adjust for your own portion.
And the app we're building follows one rule: every number comes with its range. "182 kcal (140–237)" looks less tidy than "182 kcal", but it's honest. How we calculate.
Frequently asked questions
Are photo calorie apps accurate?
They recognize foods well but estimate amounts less accurately. In a 2026 NIH test, four popular apps came in 250 to 345 calories short per meal, mostly because of fat that doesn't show up in a photo.
Is Cal AI accurate?
Cal AI was one of the four apps in the NIH study presented in July 2026. Together, the four underestimated meals by 250 to 345 calories. The study is preliminary and hasn't published results app by app.
Why do calorie apps underestimate?
A photo can't see cooking oil, butter, dressings or fat inside foods. Fat has 9 calories per gram, so a few unseen grams add up fast.
Which calorie counting app is most accurate?
None is exact from a photo. The most reliable are the ones that let you correct portions and add ingredients, and that show a margin of error instead of a single number.
Sources
- ScienceDaily (July 2026). NIDDK study presented at NUTRITION 2026, American Society for Nutrition.
- HealthDay (July 2026). How Accurate Are Photo-Based Calorie Apps? 4 Are Put To The Test.
- Li et al. (2024). Evaluating the Quality and Comparative Validity of Manual Food Logging and Artificial Intelligence-Enabled Food Image Recognition in Apps for Nutrition Care. Nutrients.
More guides
Micela shows general information. It is not medical or nutritional advice.