How accurate are photo calorie apps, and when do they fail?

What does a photo calorie app actually estimate?

It identifies the foods it can see, guesses their portion sizes from the image, and looks up nutrition figures for each. The output is the sum of several estimates, each carrying its own error.

Two of those 3 steps are hard. Identification is largely solved for common foods, and portion estimation from a flat image is not, because depth has to be inferred. Nutrition lookup then assumes a generic version of the food, which may or may not resemble what you ate.

The result is a figure that is usefully close for some meals and substantially out for others, and the pattern of which is which is predictable enough to work with.

What can a camera not see?

Five things, and they are consistent across every app using this method, because they are properties of photographing food rather than flaws in a particular product.

1. 1Cooking fat. A tablespoon of oil is roughly 120 calories and leaves no visible trace. Two tablespoons across a pan for 2 people is 120 per portion, invisible.
2. 2Sugar and fat in sauces. A glossy sauce could be reduced stock or double cream, and they differ by several hundred calories.
3. 3Portion depth. From above, 150 g of rice and 250 g of rice look nearly identical.
4. 4Anything buried. A bake or a curry shows only its surface, so the cheese under the top layer is a guess.
5. 5Which version it was. Full-fat or reduced-fat yoghurt, thigh or breast, white or wholemeal, all look similar and differ materially.

Cooking fat is the largest single source of drift, and it goes in one direction: estimates come out low. That matters more than random error, because a consistent underestimate makes a week look better than it was.

Which meals estimate well?

Separation is the variable. A plate where each component is visible and identifiable estimates well; anything mixed, layered or sauced does not.

Baking is the extreme case and worth flagging separately. A slice of cake could plausibly be anywhere within a range of several hundred calories depending on the butter and sugar in the batter, and none of that is visible in the finished slice.

How do you get firmer numbers?

For anything you cooked yourself, take the numbers from the recipe. It sounds like more work and is usually less, because a recipe is a measurement that already happened.

Total the ingredients once, divide by the yield, and you have a per-serving figure you can reuse every time you cook that dish. Cook 15 dinners on rotation and 15 calculations cover most of your year, which is the approach in calculate calories in a recipe.

• Weigh the fat you add. It is the ingredient most often left out and the one that moves the number most.
• Record the yield honestly. A recipe that says 4 servings and feeds 3 is a 33% error on every portion.
• Keep the figure with the recipe so you calculate once rather than every time.
• Use a database or barcode for restaurant and packaged food, where no recipe exists.

How wrong can an estimate be in practice?

Worth putting numbers on, because the abstract answer of approximate does not help you decide whether to trust one.

Take a homemade chicken curry with rice, a fair test since it is a common weeknight dinner and a hard case for a camera. The visible components are chicken, sauce and rice. The invisible ones are 2 tablespoons of oil used to fry the base, 400 ml of coconut milk in a pan serving 4, and however much rice is actually under the sauce.

For a dinner like that, a photo estimate landing 15% to 25% below the real figure would not be surprising. If you are eating to a maintenance level and want a rough sense of the week, that is tolerable. If you are working to a deficit that depends on the numbers, it is the difference between progress and confusion.

When is a photo estimate the right tool?

More often than the accuracy discussion suggests. A rough number you capture beats a precise one you abandon after 5 days, and consistency matters more than precision for spotting a trend.

It is the right tool when you eat a lot of food you did not cook, when you want a directional sense of a week rather than an audit, or when you have tried database logging before and stopped. It is the wrong tool when a specific target has to be hit, since a consistent underestimate on cooking fat is exactly the error a target cannot absorb.

The stronger move for most people is upstream: decide the week's meals in advance, so the food is already appropriate and the measurement is a check rather than a verdict. Meal plan for weight loss covers that side, and track macros in recipes covers getting numbers without a full diary.

It is also worth knowing that the estimate is not the only number carrying uncertainty. Nutrition figures for the same food vary between databases, portion sizes vary between people, and a recipe cooked twice is never quite identical. Nobody is working with exact figures, which argues for treating any single day as noise and any consistent 2-week direction as information.

On apps, ReciBites vs Cal AI covers photo estimation directly and ReciBites vs MyFitnessPal covers database logging.

Portion estimation

Inferring how much food is present from an image. The largest source of error, because depth is not visible from above.

Systematic error

An error that leans consistently one way, such as omitting cooking oil, which makes a week look better than it was.

Per-serving figure

The nutrition of one portion, calculated by totalling a recipe's ingredients and dividing by its yield.