Roughly, yes, but not precisely from the photo alone. In a 2025 study, a general-purpose AI model estimated the energy in 195 dishes from their photos and missed by about 123 kcal on average; when it was also given the full ingredient list with amounts, the miss fell to about 53 kcal1.
So a photo gives a starting estimate. The details you add, especially the fat that the camera cannot see, do most of the work. This article walks through what was tested, where the errors cluster, and how the newest studies compare.
What "accurate" means here
Studies of calorie estimation compare the AI's number with a reference value: the dish's true energy, from a nutritionist who weighed every ingredient, or from the recipe1. The gap is summarised in a few ways:
| Term | Plain meaning |
|---|---|
| Mean absolute error (MAE) | The average size of the miss, over or under, in kcal |
| Mean absolute percentage error (MAPE) | The same miss as a share of the dish's true value |
| Macros | Protein, carbohydrate and fat |
| General-purpose AI model | A chat assistant that reads text and images, not built for food |
A 100 kcal miss on a 400 kcal dish is a 25% error. On an 800 kcal dish it is 12.5%. Both measures matter.
In plain terms: the studies ask how far the AI's number lands from the true one, on average, dish by dish.
The 2025 study: four ways of asking
A 2025 study tested one general-purpose AI model on 195 dishes1. The dishes came from three sources: 96 recipe photos from a recipe website, 74 photos from a US research database of meals with known portions, and 25 home-cooked meals weighed and annotated by nutritionists, each photographed with a ruler for scale. Each dish was run once, in a fresh chat, with default settings.
Every dish was estimated four ways1:
- Photo only.
- Photo plus hidden details: the type and amount of fat, the type and amount of sweetener, the fat content of dairy, and the type of meat.
- Photo plus the full ingredient list with amounts.
- The ingredient list with amounts, but no photo.
What the study found
The miss shrank with every piece of information the camera could not see1:
| Case | Average miss (MAE) | As a share of the dish |
|---|---|---|
| Photo only | 123 kcal | about 31% |
| Photo + hidden details | 92 kcal | about 24% |
| Photo + full ingredient list | 53 kcal | about 14% |
| Ingredient list, no photo | 67 kcal | about 18% |
Average miss in energy by how much the AI was told · kcal
Three details stand out.
The photo still helped. With the full ingredient list but no photo, the miss grew from 53 to 67 kcal. Reading its answers, the authors found that without the image the model was sometimes unsure whether ingredients were raw or cooked1. Weight changes a lot with cooking, so that uncertainty matters.
Fat was the hardest to see. From the photo alone, the estimate for fat was off by about 53% of the true value on average, against about 43% for protein and 42% for carbohydrate1. Oil, butter and the fat in a cut of meat are mostly invisible.
Average miss by macro from the photo alone · % of true value
Big plates and bowls missed most. In the quarter of dishes with the most energy, above about 534 kcal, the photo-only miss averaged 225 kcal1. Dishes served in bowls gave the largest outliers in every source, which the authors link to how hard depth is to judge from a single flat image.
The error was also smallest on the 25 home meals that nutritionists had weighed: 75 kcal from the photo alone, 52 with hidden details and 29 with the full list. The authors suggest part of the error on the web photos came from imperfect reference values rather than from the model1.
In plain terms: a photo gets you into the right range. Telling the AI about the oil, the cut of meat and the amounts gets you much closer.
Other studies point the same way
Portion size bias. A 2025 study gave three general-purpose models 52 standardised food photos in small, medium and large portions, weighed as reference2. The two better models missed energy by about 36% on average; the third missed various nutrients by 64% to 110%. All three under-estimated more as portions got bigger. The authors judged the accuracy comparable to people's own self-reports, but not precise enough for clinical or athletic use.
Hidden fat, the other way. A 2026 study photographed 15 standard hospital meals under controlled light and compared 10 AI models with 10 registered dietitians against weighed values3. For energy and carbohydrate, the dietitians and the best models were often within 10%. For protein and fat, every AI model was less accurate, and for fat every model over-estimated, by more than 20% on average. The authors call fat an "invisible nutrient".
A different cuisine. A 2026 study tested open AI models of three sizes on 510 lunch photos from a Peruvian cookbook, with the meal name given4. The best model's calorie estimate missed by 108 kcal on average and was within the study's agreement range for 45% of meals. The authors concluded that performance "falls short of the precision required" for clinical or consumer-facing use.
In plain terms: across studies, AI photo estimates are in the right neighbourhood, weakest on fat and on large portions, and better with more context.
What this does and does not mean
It does not mean AI photo logging is useless. The 2025 portion study judged the better models comparable in accuracy with traditional self-reported methods, such as food diaries, without the effort of writing everything down2. For how large those self-report gaps are, see how accurately do you need to track calories.
It does not mean every estimate is off by 123 kcal. That is an average. Some dishes were close; large, mixed and bowl-served dishes were far off.
It does mean the photo is a draft. The studies that improved accuracy did so by adding information: the fat, the sweetener, the cut of meat, the amounts. That is description of what they tested, not advice any study gave for an app.
Models change quickly. Each study tested specific versions at one point in time. Results for newer versions may differ, in either direction.
How this looks in Zyra Nutrition
Zyra Nutrition, live at nutrition.zyra.fit, has a photo import. It returns the dish as a list of ingredients with gram amounts, each matched to its food database with a confidence badge, and has a notes field for anything the photo does not show. Every line and every gram amount can be edited before the meal is logged, and the app labels the portions as estimates from the photo. We have not measured its accuracy, so we make no claim about it. The studies above describe why the notes and the gram check exist. For how day-to-day logging errors wash out over weeks, see how apps estimate energy expenditure.
What the studies found
| Study | Who was studied | Key result | Year |
|---|---|---|---|
| Rodríguez-Jiménez et al., *Nutrients*1 | 195 dishes, one general-purpose AI model, four levels of context | Average miss 123 kcal from photo only, 92 with hidden details, 53 with full recipe | 2025 |
| Fridolfsson et al., *Curr Dev Nutr*2 | 52 standardised food photos, three general-purpose models | Energy missed by about 36% (best two); under-estimation grew with portion size | 2025 |
| Isobe et al., *Nutrients*3 | 15 hospital meals; 10 AI models vs 10 dietitians | Energy often within 10% for the best; fat over-estimated by more than 20% by all AI models | 2026 |
| Carrillo-Larco et al., *J Nutr*4 | 510 Peruvian lunch photos, open models of three sizes | Best calorie miss 108 kcal; agreement for 45% of meals | 2026 |
Limits worth knowing
- Food, not people. These studies score dishes; none tested whether AI logging changes what people eat or weigh.
- One run per dish in the 2025 study, so the model's run-to-run variation was not measured1.
- Imperfect reference values. Recipe-site and database values have their own errors, which inflate apparent AI error.
- Specific models, specific dates. Every study tested particular versions; these change often.
- Small or narrow samples. Fifteen hospital meals; 52 standard photos; one cuisine in the Peruvian study.
- Abstracts only for the 2025 portion study and both 2026 studies. The 2025 four-case study was read in full.
- Interest. The authors of the 2025 four-case study built a customised assistant for food photos, described in an appendix1.
Latest research
The newest studies we opened are from 2026. One compared AI models with registered dietitians on weighed hospital meals and found energy close for the best models but a consistent over-estimate of fat3. The other, on Peruvian dishes, judged current models not yet precise enough for consumer apps4. Neither overturns the 2025 pattern: context helps, and fat is the weak spot.
The short version
- AI can estimate calories from a photo, roughly: about 123 kcal off on average in the largest test.
- Telling it what the camera cannot see, such as fat, sweetener, cut of meat and amounts, cut the miss to about 53 kcal.
- Fat and large portions are where estimates go most wrong.
- A photo estimate is a draft to check, not a measurement.
Educational overview only — not medical advice.
