How far is a photo meal log from a weighed record?

A photo of a meal leaves a record. Whether that number matches food on a scale is a different question. A review of photo records, and a feeding trial that weighed what people ate without announcing it, found errors of different sizes.

What the studies looked at

Dang Khanh Ngan Ho and colleagues, in a 2020 systematic review, pooled photo-based ways of recording meals against energy and macronutrients. The comparisons were doubly labeled water, 24-hour recalls, and estimated or weighed food records. Of 4,346 papers, 13 studies and 606 participants remained.

Clare Whitton and colleagues ran a different design in 2024. One hundred fifty-two people ate breakfast, lunch, and dinner in a research facility, and the foods and drinks they consumed were weighed discreetly. The next day, intake was estimated with one of four methods. For the mobile Food Record, a trained analyst read the meal photos (mFR-TA). Another method was an interviewer-led 24-hour recall that used those photos (IA-24HR).

Results

In Ho’s pooled estimate, image-based methods under-reported energy by 179.32 kcal (95% CI −269.50 to −89.15). The studies disagreed strongly (I² = 89%). Against doubly labeled water the gap was larger, −448.04 kcal (95% CI −755.52 to −140.56).

Protein (−0.08 g) and fat (−0.57 g) were not statistically different. Carbohydrate only tended to be lower, by 9.17 g. Setting doubly labeled water aside, image methods were not statistically different from recalls and food records. The authors wrote that image methods carry serious measurement error, as traditional methods do.

Averages were closer in Whitton’s controlled meals. The mean difference between true and estimated energy, as a percentage of true intake, was 1.3% for mFR-TA (95% CI −1.1 to 3.8) and 15.0% for IA-24HR (11.6 to 18.3). In the same trial Intake24 was 1.7% and ASA24 was 5.4%. Only Intake24 also matched the distribution of energy and protein, not just the average.

Limits

179 kcal is an average across studies with different designs. Apps, analysts, days of recording, and participants all differed, so that number should not be added to or subtracted from one person’s lunch today.

The Whitton trial was one day, with food that was served, and with a person reading the photos. That is not an app reading a home-cooked bowl, a delivery meal, or a sauce hidden under the rice. An average difference of 1.3% does not mean every meal landed within 1.3%.

What this means for a log

Photos are good at leaving a trace of what was eaten, and weak where oil, drinks, and rice underneath cannot be seen. In the free-living research, those gaps usually pull energy down. If a record looks wrong, edit the amount next to the photo. Read the daily total as the record you kept, not as a scale weight.