How accurate is photo logging?
In short
Photo logging is useful, but not exact. In studies, logging from images tends to come out lower than what people really eat, mostly because portions are hard to judge from a picture. That is also true for writing food down by hand. In Mitta, the photo only suggests foods and portions; the nutrition values come from food databases, and you check every item before anything is logged.
Cody de LeeuwLast checked 29 October 2026 · 5 min read
What the research shows
A meta-analysis of 13 studies compared photo-based food records with other ways of measuring intake. On average, the photos came out about 180 kcal a day lower than the comparison methods, and about 450 kcal lower than the most accurate method, which measures energy use with labelled water.1
That does not make photos worse than the alternative. When people write down what they eat, they also report less than they really eat. In one well-known study, people who struggled to lose weight underreported their food intake by almost half.2
Where the errors come from
- Portion size. A photo shows how big a portion looks, not how much it weighs. This is the biggest source of error.
- Hidden ingredients. Oil, butter, sauces and sugar in a dish are hard or impossible to see.
- Similar-looking food. Full-fat and low-fat yoghurt, or white and wholemeal pasta, look the same on a plate.
How to get the most out of it
- Take the photo from above, with the whole plate in view.
- Check the portions. Change a grams figure if you know better, for example from the pack.
- Add what you cannot see: the oil you cooked with, the dressing, the sugar in your coffee.
- For foods you eat often, a kitchen scale once or twice teaches you what a portion really weighs.
What counts over weeks is your average. Small errors on single meals matter less than logging most of what you eat.
Moderate evidence
A meta-analysis of 13 studies found that image-based methods underestimate intake on average, with large differences between studies. Most of those studies were not about AI apps specifically. How we rate evidence
How Mitta does it
The photo suggests. You decide.
In Mitta, a photo of your plate gives you a list of foods with portions to check. Nothing is logged until you approve it.
- Values from databasesThe AI only recognises foods and estimates grams. The calories and protein come from food databases such as NEVO, CoFID and USDA, not from the AI.
- You check every itemChange a portion, swap a food or remove it before you save the meal.
- Pro + AISnap your plate is part of Pro + AI, with a fair-use limit of 400 AI food logs a month.
Log a meal in seconds, and check it.
Snap your plate, check the portions, save. Free to start, in the web app for invited testers.
Questions about
How accurate is photo logging
Is photo calorie counting accurate?
It is a useful estimate, not a measurement. Research finds image-based logging tends to come out lower than real intake, mostly because portions are hard to judge. Checking the portions helps.
Is it less accurate than a food diary?
Not necessarily. People who write their food down also tend to report less than they eat. Both work best when you check portions and log consistently.
Does the AI decide the calories?
Not in Mitta. The AI recognises foods and estimates grams; the nutrition values come from food databases, and you approve every item.
Where does my photo go?
Only when you take it, the photo goes to Anthropic, the AI provider, to recognise the food. Anthropic does not train models on it and in principle keeps it for at most 30 days. Your food log stores the foods and grams, not the photo.
Sources
- Ho DKN, Tseng SH, Wu MC, Shih CK, Atika AP, Chen YC, Chang JS. Validity of image-based dietary assessment methods: a systematic review and meta-analysis. Clinical Nutrition, 2020; 39(10): 2945–2959. doi.org/10.1016/j.clnu.2020.08.002
- Lichtman SW, Pisarska K, Berman ER, Pestone M, Dowling H, Offenbacher E, et al. Discrepancy between self-reported and actual caloric intake and exercise in obese subjects. New England Journal of Medicine, 1992; 327(27): 1893–1898. doi.org/10.1056/NEJM199212313272701
Written by Cody de Leeuw. First published 29 October 2026, last checked 29 October 2026. This article is general information, not medical advice. Spotted a mistake? Tell me