A calorie estimate from one food photo is close enough to show the shape of your eating and too rough to trust to the last 100 kcal. In a 2025 test on standardized meal photos, the best AI models’ energy estimates were off by about 36% on average [1]. Food diaries and recalls carry errors of a similar size, and they mostly err on the low side [2].

What “accurate” means for a food log

Two kinds of error matter.

Error per meal. How far one estimate lands from the true value, in either direction. Studies often report it as the mean absolute percentage error (MAPE).
Bias. Whether estimates lean consistently high or low. Random errors in both directions partly cancel over a week of meals. A consistent bias adds up.

The reference matters too. Photo methods are tested against weighed food, which checks what was on the plate, or against doubly labeled water, a reference method that measures the energy a person spends in daily life. When body weight is stable, intake and expenditure are about equal, so doubly labeled water shows how much a person really ate.

How accurate other methods are

Self-reported intake is the usual baseline, and it tends to come in low. A systematic review of 59 studies with 6,298 adults compared self-reported energy intake with doubly labeled water [2]. Most studies found significant under-reporting:

Method	Under-reporting vs doubly labeled water
24-hour recall	8–30%
Food record (diary)	11–41%
Food frequency questionnaire	4.6–42%
Smartphone image-based methods	20–37%

Of these, 24-hour recalls showed the smallest and most consistent under-reporting [2].

Photos can come close on average. In a validation of the Remote Food Photography Method, people photographed their food and their leftovers with a smartphone and sent the images for analysis. Estimated intake in daily life did not differ significantly from doubly labeled water: the mean difference was −152 kcal per day, with a standard deviation of 694 kcal [3]. The average was close; individual results varied widely. Portion estimates from photos can match in-person judgment: in an earlier cafeteria study, observers’ estimates from digital photos correlated highly with weighed portions and were about as accurate as estimating by eye at the table [4].

A 2020 review of image-assisted methods concluded that less burdensome methods tend to be less accurate, that no current method is adequate in every setting, and that fully automated intake assessment with acceptable precision was not yet a reality [5].

How accurate AI photo estimates are

Vision-capable language models have since taken over much of the automated part. In the 2025 study, researchers gave 52 standardized food photos to ChatGPT-4o, Claude 3.5 Sonnet and Gemini 1.5 Pro: single foods and complete meals, each in three portion sizes [1].

ChatGPT and Claude estimated weight with a MAPE of 36–37% and energy with a MAPE of 35.8%.
Gemini’s errors were larger, with MAPE values of 64–110%.
Estimates from ChatGPT and Claude correlated with the true values at 0.65–0.81.
All three models underestimated more as portions grew larger.

The authors concluded that the better models matched the accuracy of traditional self-report without its effort for the user, and that they are unsuited to precise clinical or athletic dietary assessment [1]. Model versions change quickly, so a result for one version needs retesting for the next.

Where the errors come from

Source of error	Why a photo struggles	What helps
Portion size	Depth, bowl volume and plate size are hard to judge from one angle	A second angle, a known object in the frame, weighing foods you eat often
Hidden fats	Cooking oil, butter and dressing are mostly invisible	A sentence that names them
Mixed dishes and sauces	Ingredients and their ratios sit out of sight	Saving the recipe once as your own food
Drinks	A glass looks the same with or without sugar or cream	Naming the drink, or scanning the product
Composition data	A database value describes a typical food, not your version	Barcodes and labels for packaged food

Labels have their own tolerance. Under US rules, a food is misbranded if its calories exceed the declared value by more than 20% [6]. In a laboratory analysis of 24 US snack foods, measured metabolizable energy averaged 4.3% more than the label stated, well within that limit [7].

Put together, a photo estimate is best read as a range. The most useful question for any single meal is which one or two items carry most of the uncertainty.

How to get better estimates

Photograph the meal before you start, from above and at an angle, with the whole plate in the frame.
Keep a fork, a hand or a standard glass in the picture as a size reference.
Photograph what you leave, or note it.
Add one sentence for what the photo can’t show: “fried in a tablespoon of butter,” “two teaspoons of sugar in the coffee.”
Scan the barcode of packaged foods.
Weigh a food you eat often once, and reuse that portion afterwards.
Check the largest and most uncertain items first. A 10% error on the main dish outweighs a 50% error on a garnish.
Judge your intake from a week of logs. Single-meal errors partly cancel over that span.

If you use calorie or carbohydrate numbers for a medical purpose, such as insulin dosing, confirm portions by weighing or from the label. Diagnosis and treatment decisions belong with you and your clinician.

How MyMeals handles this

MyMeals refuses to fake precision. A frontier vision model describes each item and its portion cues, then MyMeals matches every item to food-composition data from USDA FoodData Central, Fineli and Open Food Facts. The nutrient values come from that data, not from the AI model. A barcode or a photographed label identifies a product exactly and takes precedence. For uncertain items, MyMeals calculates an honest range, such as 170–330 kcal for two slices of bread with cheese, and flags the item to calibrate. MyMeals calibrates to your staple foods, brands and portions through your corrections, grocery receipts and Personal Tableware, so the ranges narrow as you go. The science page details each step.

Sources

Fridolfsson J, Sjöberg E, Thiwång M, Pettersson S. Performance evaluation of 3 large language models for nutritional content estimation from food images. Current Developments in Nutrition. 2025;9(10):107556. https://doi.org/10.1016/j.cdnut.2025.107556
Burrows TL, Ho YY, Rollo ME, Collins CE. Validity of dietary assessment methods when compared to the method of doubly labeled water: a systematic review in adults. Frontiers in Endocrinology. 2019;10:850. https://doi.org/10.3389/fendo.2019.00850
Martin CK, Correa JB, Han H, Allen HR, Rood JC, Champagne CM, Gunturk BK, Bray GA. Validity of the Remote Food Photography Method (RFPM) for estimating energy and nutrient intake in near real-time. Obesity. 2012. https://doi.org/10.1038/oby.2011.344
Williamson DA, Allen HR, Martin PD, Alfonso AJ, Gerald B, Hunt A. Comparison of digital photography to weighed and visual estimation of portion sizes. Journal of the American Dietetic Association. 2003. https://doi.org/10.1016/s0002-8223(03)00974-x
Höchsmann C, Martin CK. Review of the validity and feasibility of image-assisted methods for dietary assessment. International Journal of Obesity. 2020. https://doi.org/10.1038/s41366-020-00693-2
US Code of Federal Regulations. 21 CFR 101.9(g)(5), Nutrition labeling of food. Legal Information Institute, Cornell Law School. https://www.law.cornell.edu/cfr/text/21/101.9
Jumpertz R, Venti CA, Le DS, Michaels J, Parrington S, Krakoff J, Votruba S. Food label accuracy of common snack foods. Obesity. 2013. https://doi.org/10.1002/oby.20185

How accurate are calorie estimates from food photos?