How accurate are photo calorie apps?

What did the 2026 NIH-linked test of photo calorie apps find?

The most useful recent evidence comes from researchers in the intramural program of the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK). Olivia Charles presented the work at NUTRITION 2026, the American Society for Nutrition's annual meeting, on July 25, 2026.

The setup is what makes it worth reading. The team used 102 standardized meal photos from an NIH inpatient diet study, where every ingredient had been weighed to 0.1 g in a metabolic kitchen. That gives a true calorie count to compare against, which most app reviews never have.

They ran the photos through four apps: MyFitnessPal, Lose It!, Cal AI and Appediet. The findings, as reported by the press release and HealthDay:

• All four apps underestimated energy, by roughly 250 to 345 calories per meal on average.
• Fat was undercounted by about 30 g per meal on average.
• Carbohydrate estimates were more consistent than fat and protein estimates.
• MyFitnessPal and Lose It! were more accurate for higher-calorie meals than for lower-calorie ones.
• In an exploratory follow-up of more than 200 ketogenic-diet meals, accuracy was worse, likely because fat was consistently underestimated.

"These apps tend to underestimate calories, especially from fats," said Aaron Hengist, PhD, a postdoctoral fellow at NIDDK, in the release.

Even with those limits, the direction is clear and matches earlier research: photo apps tend to report less food energy than is on the plate.

Is AI calorie counting accurate in peer-reviewed research?

Three published studies are worth knowing, and they point the same way.

Recognizing food is the easy part

A 2024 study in Nutrients by Li and colleagues tested seven apps with AI image recognition. MyFitnessPal (97%) and Fastic (92%) were best at identifying the food components in a photo. But the authors concluded that "automatic energy estimations from AI-enabled food image recognition were inaccurate." For Asian mixed dishes, energy was underestimated by up to 76%.

That gap matters. An app can name every food on your plate correctly and still get the calories badly wrong, because naming the food is a different problem from judging how much of it there is and what it was cooked in.

Error ranges are wide

A 2023 systematic review in Annals of Medicine (Shonkoff et al.) pulled together 52 studies of AI image-based dietary assessment. The studies were too different to combine in a meta-analysis. Mean relative calorie error ranged from 0.10% to 38.3%, and results were better for simple foods. The authors found AI comparable to human estimators, with potential to do better.

General chatbots miss too

A 2025 peer-reviewed study in Current Developments in Nutrition (Fridolfsson et al.) tested three large language models on 52 standardized photos against weighed food. The mean absolute percentage error for energy was 35.8% for GPT-4o, 35.8% for Claude 3.5 Sonnet and 64.2% for Gemini 1.5 Pro, which overestimated. All three models underestimated more as portions grew. If you want to try a chatbot instead of an app, our guide on whether you can count calories from a photo with ChatGPT covers the trade-offs.

Is manual calorie logging any more accurate?

Not reliably, and this is the part most coverage of the NIH test leaves out. Self-reported food intake has a long record of running low.

• A 1992 New England Journal of Medicine study (Lichtman et al.) found a small group of 10 "diet-resistant" participants under-reported intake by 47% on average and over-reported activity by 51%.
• A 1995 review by Schoeller noted that individual underestimates of 50% are "not uncommon," with about 20% individual error even in non-obese adults.
• A 2025 Nature Food paper (Bajunaid et al.) built an equation from 6,497 doubly labelled water measurements and applied it to UK and US national surveys. It flagged misreporting in more than half of the reports.

Manual logging apps are not immune either. In the same Nutrients study, manual-logging apps overestimated Western diets by 1,040 kJ (about 250 kcal) and underestimated Asian diets by 1,520 kJ (about 360 kcal).

So the honest comparison is not "photo apps are inaccurate, manual logging is accurate." Both methods miss. They miss in different ways, and photo logging is faster, which helps people keep logging. We compare the two in detail in photo logging vs manual calorie counting.

Where do photo calorie errors come from?

The research points to a few repeat offenders. Knowing them tells you where to look when you check an estimate.

• Hidden fat. Cooking oil, butter, dressings and creamy sauces add a lot of energy and are hard to see. In the NIH test, fat was the nutrient most often missed, and high-fat keto meals did worse.
• Portion size and depth. A single photo is flat. It shows the top of a bowl, not how deep it is or how densely food is packed. Errors grew with portion size in the 2025 chatbot study.
• Mixed dishes. Stews, curries, stir-fries and rice dishes hide their ingredients. That is where the Nutrients study saw underestimates of up to 76%.
• Small or lower-calorie meals. Two of the NIH-tested apps were less accurate on lower-calorie meals, where a fixed error is a bigger share of the total.

We go deeper on each cause in why AI photo apps underestimate calories.

Why can't you compare vendor accuracy percentages?

Many apps publish their own accuracy figures. They are rarely measured the same way, so putting them side by side tells you little.

When you read an accuracy claim, ask four questions:

1. What was measured? Identifying foods, total calories or each macro are very different tasks.
2. Against what? Weighed food in a lab is the gold standard. Comparison against another app or a database is not.
3. Which meals? Simple single foods are much easier than mixed, oily dishes.
4. Who ran it? A company testing its own app on its own meals is not an independent result.

Our page on the most accurate AI calorie tracker apps goes through this in more detail.

How do you get better numbers from a photo calorie app?

You cannot make a photo estimate perfect, but you can remove most of the predictable error. The research tells you exactly where to spend your attention.

1. Check the fat first. If the food was fried, sautéed, dressed or buttered, assume the estimate is low and raise it. This is the single biggest known gap.
2. Correct the portion. Compare the app's portion estimate with what you actually ate. If you served yourself a second helping or left half, adjust it. Our guide to estimating portion sizes without a scale helps here.
3. Take a better photo. Shoot from a slight angle so depth is visible, use good light, and keep the whole plate in frame. See our food photo tips for accurate calorie tracking.
4. Use the barcode for packaged food. Label data beats a visual guess every time it is available.
5. Add what the camera can't see. A drink, a sauce on the side, sugar in coffee: log it or add it to the numbers.
6. Look at trends, not single meals. If your log runs consistently low, the trend over weeks still tells you something useful.

Where does FitFoodie fit in?

We make FitFoodie, a food photo nutrition tracker, so here is how it relates to the evidence above.

FitFoodie was not in the NIH test. It uses Google Gemini to estimate a meal from one photo. We found no peer-reviewed weighed-meal test of the current Gemini models, and the 2025 study above tested an older one, so we do not publish an accuracy figure for FitFoodie. Its results are estimates, and the app is built around checking them.

• Review before saving. The result is one estimate for the whole dish. You can scale the portion from 0.5× to 2× and edit all eight values: calories, protein, carbs, fat, fiber, sugar, sodium and saturated fat.
• Ingredients are a checklist, not a calculator. You can add or remove ingredient names, but that does not change the numbers. If you spot missing oil, raise the fat and calories yourself.
• Barcode data when it exists. If your photo shows a product barcode, FitFoodie uses the label data instead of guessing. The barcode scanner is free and checks Open Food Facts, USDA FoodData Central and FatSecret in turn.

AI photo and text analysis are part of FitFoodie Premium. Barcode scanning, water and weight tracking are free.

Nutrition values from any app are estimates for information, not medical advice. If you are managing a health condition, work with a registered dietitian or your doctor.

Frequently asked questions