On this page
Key takeaways
- Photos estimate the visible. Identification of plated foods is good; portion size is the main source of error, and hidden fats and sugars are the main source of underestimation.
- People are worse. In a classic study, adults unknowingly under-reported their intake by 47 percent and over-reported activity by 51 percent.1
- Research-grade photo methods have come within about 5 to 7 percent of true intake in validation studies, far closer than self-report.2, 3
- Labels are not exact either. US rules allow calories to run 20 percent above the declared value and still comply.4
- Consistency beats precision. A steady bias disappears when your targets are set from your weight trend rather than from the raw log.6
What a photo can and cannot see
A food photo carries three kinds of information: what the food is, roughly how much of it there is, and clues about how it was prepared (grilled versus fried, dressed versus plain). Modern image models are good at the first, fair at the second and limited on the third. Four things cause most of the error:
- Portion scale. A photo is two-dimensional. Plate size, camera angle and distance all change how big a serving looks, and depth (how deep the bowl is, how many layers in the lasagna) is mostly guessed.
- Density. A cup of salad and a cup of rice look similar in volume and differ five-fold in calories. Visual volume has to be converted to weight, and the conversion is a guess per food.
- Hidden ingredients. Oil absorbed during frying, butter melted into vegetables, cream in a soup, sugar in a glaze, dressing already tossed through. These are the single biggest cause of underestimation, and they are invisible by definition.
- Mixed dishes. A curry, a casserole or a burrito hides its proportions. The model sees the surface and infers the rest.
Put together, photo estimates tend to be good for simple plated foods (a chicken breast, a bowl of oats, a piece of fruit), weaker for restaurant and mixed dishes, and systematically low where hidden fat is involved.
The human baseline
Judge the technology against the alternative, which is people writing down what they ate. In a well-known study in the New England Journal of Medicine, adults who were certain they could not lose weight on their diet were measured with precise methods and found to be under-reporting their intake by an average of 47 percent, while over-reporting their physical activity by 51 percent. They were not lying; they genuinely could not see it.1 Even researchers who defend self-report dietary data acknowledge the measurement error and argue for using the data with that error in mind rather than discarding it.7
A photo removes two of the biggest human failure modes: forgetting an item entirely, and misjudging a portion from memory hours later. What it cannot remove is the hidden-ingredient problem, which humans also miss.
What the research says
Photo-based dietary assessment has been studied for well over a decade, mostly in research settings where trained analysts or software estimate intake from before-and-after images. In the validation of one such method, the Remote Food Photography Method, estimated energy intake in free-living adults came within roughly 5 to 7 percent of the true value, a fraction of the error typical of self-report.2 A review of image-assisted and image-based methods concluded they reduce the burden on the user and improve accuracy relative to traditional recall, while noting that portion estimation and mixed dishes remain the hard problems and that performance varies by food type.3
Consumer apps that estimate from a single photo, without a reference object or a second "leftovers" photo, cannot be expected to match research-grade methods. The honest expectation is a useful ballpark that is right about the food, approximately right about the amount, and low on hidden fat, with errors that are largest for the dishes you would struggle to estimate yourself.
Labels and databases are estimates too
It is tempting to treat the nutrition label as truth and the photo as a guess. Both are estimates. US regulations consider a label compliant as long as measured calories, sugars, fat and sodium do not exceed the declared value by more than 20 percent, and rounding rules add more slack on top.4 Database values, including the USDA's own, are averages across samples of a food, and a home-cooked or restaurant version of the same dish can differ from the database entry by the amount of oil in the pan.5
None of this makes labels useless. It makes the standard clear: a logging method is good enough when it is consistent and close, not when it is exact, because exact is not available anywhere.
Which method, when
| Food | Best method | Why |
|---|---|---|
| Packaged food with a barcode | Barcode scan | Pulls the manufacturer's values from a database; the serving count is the only thing you have to judge. |
| Packaged food without a barcode match | Nutrition label photo | Reads the panel as printed, serving size included. |
| Food you cooked and can weigh | Manual entry with a kitchen scale | Weight removes the portion error; the database value is the only estimate left. |
| Restaurant meals, meals out, anything you cannot weigh | Photo, then adjust | Better than memory; add the oil, dressing or sauce you know is there. |
| Drinks | Barcode or manual | Liquids are hard to estimate visually and often carry most of a meal's sugar. |
Why the weight trend fixes it
Here is the part most tracking advice misses. If your logging method runs 15 percent low every day, your log is wrong but your trend is still useful, because the error is consistent. Researchers have shown that long-term energy intake can be measured accurately from repeated body-weight measurements alone, using the known relationship between energy balance and weight change, without relying on what people report eating.6
A tracking plan that uses this idea sets your targets from what your weight actually did against what you logged, rather than from a formula. If you log 1,800 a day and lose nothing over four weeks, the plan concludes your true burn is about 1,800 at whatever bias your logging has, and lowers the target. The systematic error in the photos, the labels and the database cancels out, because you are being measured in the same units you are logging in. What this cannot fix is inconsistency: logging carefully on weekdays and not at all on weekends breaks the relationship.
The tracking literature points the same way: it is the act of consistent self-monitoring, not its precision, that is reliably associated with weight-loss success.8
Getting better estimates from photos
- Shoot from above, in good light, with the whole plate in frame. Angled, dim or cropped photos are the main cause of failed or inflated estimates.
- Include a familiar object for scale when you can: a fork, a hand, a standard plate.
- Add what the camera cannot see. A tablespoon of oil is about 120 calories; a tablespoon of dressing 60 to 90. If you know it went in, log it.
- Set the number of servings honestly. If the estimate describes one portion and you ate two, say so; a half-portion left on the plate works the same way in reverse.
- Calibrate occasionally. Weigh a few of your regular meals once, compare with the photo estimate, and learn which way your meals run.
- Log every day, including the bad ones. A consistent log with a known bias is far more useful than a precise log with gaps.
Questions people ask
Can AI count calories from a picture accurately?
It can produce a useful estimate of visible food, good for simple plated items and weaker for mixed dishes and anything with hidden oil, butter, cream or sugar. Research-grade photo methods have come within about 5 to 7 percent of measured intake,2 while ordinary self-report can be off by nearly half.1 Single-photo consumer estimates sit between those, and run low on hidden fat.
Is a food scale more accurate than a photo?
For the amount, yes: weighing removes portion-size error, which is the largest error in a photo. The nutrition values per gram still come from a database average, so a weighed meal is a better estimate, not an exact one.5 For home cooking that you can weigh, use the scale; for meals out, use the photo and adjust.
Why does the app underestimate restaurant meals?
Restaurant food is cooked with more oil, butter and sugar than a photo can show, and portions are larger than they look on a wide plate. Add the fat you know is there, set the serving count honestly, and expect restaurant estimates to run low.
Does being off by 100 calories a day matter?
Less than it seems, if the error is consistent. A steady bias is absorbed when your targets are set from your weight trend rather than from the raw log, because you are measured in the same units you log in.6 Inconsistency matters far more than a fixed offset.
Is logging worth it if the estimates are imperfect?
Yes. Consistent self-monitoring is one of the behaviors most reliably associated with weight-loss success in the research, and the benefit does not depend on the numbers being exact.8 Even a label is allowed a 20 percent margin; the goal is a consistent record, not a perfect one.4
How Calobook uses this
Calobook is built around the limits described here. The photo scanner lists only the ingredients it can actually see, never guesses hidden sauces or oils, and tells you when it is unsure instead of inventing a number. Every value is editable on the review screen before you save, barcode and label scanning are there for packaged food, and a kitchen-scale entry is one tap away. Most importantly, the weekly check-in measures your real burn from your weight trend and what you logged, so a consistent bias in any method is corrected automatically rather than silently stalling your plan.
Photo, barcode, label or scale, with every value editable and a weekly check-in that corrects for the errors no estimate can avoid.
Sources
- Lichtman SW, et al. Discrepancy between self-reported and actual caloric intake and exercise in obese subjects New England Journal of Medicine, 1992.
- Martin CK, et al. A novel method to remotely measure food intake of free-living individuals in real time: the remote food photography method British Journal of Nutrition, 2009.
- Boushey CJ, et al. New mobile methods for dietary assessment: review of image-assisted and image-based dietary assessment methods Proceedings of the Nutrition Society, 2017.
- US Code of Federal Regulations. 21 CFR 101.9, Nutrition labeling of food Paragraph (g) on compliance.
- US Department of Agriculture. FoodData Central
- Sanghvi A, Redman LM, Martin CK, Ravussin E, Hall KD. Validation of an inexpensive and accurate mathematical method to measure long-term changes in free-living energy intake American Journal of Clinical Nutrition, 2015.
- Subar AF, et al. Addressing current criticism regarding the value of self-report dietary data Journal of Nutrition, 2015.
- Burke LE, Wang J, Sevick MA. Self-monitoring in weight loss: a systematic review of the literature Journal of the American Dietetic Association, 2011.
This article is general nutrition information written by the Calobook team from the sources listed above. It is not medical advice and does not replace a consultation with a doctor or registered dietitian. If you are pregnant, breastfeeding, under 18, or managing a medical condition, talk to a professional before changing your diet. Last reviewed October 17, 2026.