Answer · 2026

How accurate are AI photo calorie counting apps?

They estimate. A photo cannot show oil, sauces, hidden ingredients, or portion weight. NIH researchers presented tests of four photo apps against 102 weighed kitchen meals in July 2026. All four undercounted calories by about 250 to 345 per meal and fat by about 30 grams. Confirm foods and portions before you trust the day’s total.

Updated October 6, 2026 · Written by LAYERTWO, the makers of NutriCam. We make NutriCam. The other apps on this page are not ours. We have not run an accuracy test of our own, so this page makes no accuracy ranking.

Download NutriCam on the App Store Read what NutriCam is

What independent tests found

Test	What was on the plate	Result	Status
NIDDK at NUTRITION 2026	102 metabolic-kitchen meals, weighed to 0.1 g, photographed, then run through MyFitnessPal, Lose It!, Cal AI, and Appediet	All four undercounted calories by about 250 to 345 per meal and fat by about 30 g, roughly one third. Carbs were more consistent than fat. High-fat keto meals looked harder in a follow-up of 200+ meals	Conference abstract, July 25, 2026. Not yet peer reviewed
O’Hara et al., Nutrients, 2025	114 photographs of 38 Irish survey meals in three portion sizes, sent to ChatGPT-4	Food identification precision 93%. Small portions matched weight. Medium and large portions did not. ChatGPT underestimated weight on 87 of 114 photos. Mean absolute percentage difference in meal weight 27.8%. 13 of 16 nutrients more than 10% off	Peer reviewed
Fridolfsson et al., Current Developments in Nutrition, 2025	52 standardized photos of components and meals, three portion sizes, cutlery for scale. ChatGPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro	ChatGPT and Claude were off by 35.8% on calories on average. Gemini 64.2%. All three underestimated more as the portion got larger	Peer reviewed
Hoang et al., JAMA Network Open, 2023	222 named foods, text only, no photograph. ChatGPT-3.5 and ChatGPT-4 vs a nutritionist database	Energy within ±10% for about 35% to 48% of items, depending on model and language	Peer reviewed. Named foods, not photo logging

The kitchen test is the one that matches how people actually use these apps: a picture of a mixed plate, no typed gram weight. ASN’s own note on the abstract: treat the numbers as preliminary until a journal paper appears. The direction is the useful part. Uncorrected photo logs ran low, and fat was the nutrient that went missing.

What a photo cannot see

On the plate	Can a photo see it?	Why the calorie number moves
Oil in the pan	No	Almost all fat, almost no visual trace after cooking
Dressing, mayonnaise, sauces	Rarely	A few squirts change energy a lot and look like shine
Hidden filling (sandwich, burrito, soup)	No	The model sees the outside
Portion weight	Guess	Same recipe on two plates often gets one standard serving
Air-fried vs deep-fried	Unreliable	The surface can look similar. Calorie density does not
Brand, milk type, fortification	No	The package, not the picture, holds those numbers

People who log by camera run into the same two failures. The app names the food. It cannot weigh it. Identification is mostly solved. Volume is not. A 2D picture has no plate diameter, no density, and no view under the top layer.

That is also why a chatbot and a dedicated calorie app can look equally confident and still miss the same oil. The limit is the photograph, not the brand name on the icon.

What vendors claim, and what they measured

App	Its own accuracy line	Independent photo test
Cal AI	FAQ, last updated July 2026: “about 80% accurate.” Terms of Service: calorie estimates are approximations and may be incomplete or inaccurate	In the NIH four-app set. Undercounted with the other three when portions were not adjusted
SnapCalorie	FAQ: about 15% mean caloric error on a test set its team built. Oils, cooking fats, and sugar in restaurant food are an educated guess from averages and visual cues	Not in the NIH four-app set. The 15% figure is the company’s own evaluation, not a metabolic-kitchen retest
Cronometer Photo Log	Blog: the photo is a search. Numbers come from NCCDB and CFCD after you review ingredients and portions. It suggests missed extras such as butter, oil, and condiments	Not an accuracy ranking. The saved number is the database row you confirm, not a camera measurement
NutriCam	We have not published an accuracy percentage. The product asks you to review the estimate and change portion size before it becomes the log	We have not run a controlled test. Every number from a photo is an estimate

Vendor percentages are not interchangeable. “80% accurate” does not say 80% of what, on which meals, against which scale. SnapCalorie’s 15% is a mean error on its own dishes. Neither claim is the NIH kitchen result.

Cal AI remains the well-known fast photo counter of calories, protein, carbs, and fat. Full split: NutriCam vs Cal AI.

How to make the estimate less wrong

Treat the first number as a draft. Confirm the foods. Change the portions. Add the oil, dressing, and anything under the top layer. Packaged items belong on a barcode, not a plate photo.

Cronometer is built for that confirm step. Photo Log is Gold-only. It maps the picture onto database foods, then waits for you to swap, add, and size them. If you cook at home, its own blog says to put the plate on a kitchen scale and include the weight in the shot. That is the precision-diary job. Comparison: NutriCam vs Cronometer.

Repeated meals are the ones worth weighing once. An error on a plate you eat five times a week stacks. A one-off restaurant photo is a guess you correct by eye.

NutriCam, on iPhone, is not a more accurate camera. Photograph the meal or say what you ate, change the portion before it is the day’s record, then ask what to eat next from that log. The advisor uses today’s meals, targets, workouts, and Apple Health sleep when connected. Details: what should I eat next. Other photo apps, by job rather than by score: best AI calorie tracker apps.

No app can measure a meal from a photo. A picture cannot show oil, sauces, hidden ingredients, or portion weight. Every number from every app on this page is an estimate until you correct the foods and portions. We have not run an accuracy test across these apps, so this page makes no accuracy ranking.

FAQ

Can a photo calorie app see oil, dressing, or cooking fat?

No. Oil in the pan, salad dressing, and mayonnaise barely show in a picture. That is why the NIH kitchen tests missed about 30 grams of fat per meal. Add the oil or sauce yourself, or weigh meals you cook often. The photo is a start, not a lab assay.

Did the NIH test include Cal AI?

Yes. NIDDK researchers ran photos of 102 metabolic-kitchen meals through MyFitnessPal, Lose It!, Cal AI, and Appediet, and presented the work at NUTRITION 2026. All four undercounted calories by about 250 to 345 per meal. The findings are a conference abstract, not yet a peer-reviewed paper.

Is SnapCalorie’s 15 percent error an independent lab result?

No. SnapCalorie’s FAQ says about 15% mean caloric error on a test set its team built, including dishes with oils and occlusions. That is the company’s own evaluation, not a metabolic-kitchen retest. Independent NIH kitchen photos did not include SnapCalorie.

Are ChatGPT and Claude more accurate than dedicated calorie apps?

Not in a head-to-head on the same kitchen meals. Peer-reviewed photo tests of ChatGPT-4 and Claude put energy error around 36%, with larger plates underestimated. The NIH four-app test used a different protocol. Neither result ranks consumer apps against chatbots.

Which photo calorie app is the most accurate?

We have not run a test, so we do not rank them. The NIH kitchen photos undercounted on all four apps tested. Cronometer’s Photo Log maps the picture to a verified database after you confirm foods and portions. That is a diary workflow, not a claim that the camera measured the plate.

Are numbers from a food photo medical or lab values?

No. They are estimates. None of the apps on this page is a medical device. Oils, sauces, hidden ingredients, and portion weight can throw any photo log off. Confirm the foods before you treat the day’s total as a record.

Sources

Every fact about another app on this page comes from that app’s own site, help center, or store listing, checked October 6, 2026. Prices and features change. Check the live listing before you buy.