How accurate is AI photo calorie counting?

A multimodal model looks at your photo and identifies what is on the plate, then estimates portion size from visual cues — the size of the plate, the depth of the bowl, how the food is arranged. It maps each identified item to nutrition data and adds up the result.

Two of those steps are genuinely hard. Identification is the easier one: a chicken breast, a bowl of rice, and a side of broccoli are recognizable. Portion estimation is harder, because a photograph flattens depth, and the difference between 100 g and 160 g of rice can be a matter of a few pixels.

The third difficulty is invisible ingredients. A tablespoon of olive oil is around 120 kcal and leaves almost no visual trace on a finished dish. Cream in a sauce, butter on vegetables, and sugar in a marinade all behave the same way.

It is strongest on separated, plated food: a piece of protein, a starch, and a vegetable, photographed from a slight angle in decent light. Packaged food scanned by barcode is more accurate still, because the numbers come straight from the label rather than from an estimate.

It struggles with homogeneous mixtures — stews, curries, smoothies, casseroles — where the ingredients are not individually visible. It struggles with restaurant food, which is routinely richer than it looks. And it struggles with anything shot in a way that hides scale, such as a close-up with no reference object in frame.

This is exactly why CalEasy lets you combine a photo with a voice note or a line of text in a single log. "Cooked in two tablespoons of olive oil" supplies precisely the information a photograph cannot carry, and it takes three seconds to say.

Suppose your estimates run 10% low across the board. You set a target, track for three weeks, and the scale barely moves. You lower the target by 200 kcal and it starts moving. The bias is now absorbed — because you calibrated against your own results rather than against an absolute truth.

Now suppose your estimates are perfectly accurate but you only log four days a week. You have no idea what your intake is, and no trend to calibrate against. The second person is far worse off than the first, despite better per-meal accuracy.

This is the honest case for photo logging: it is not the most precise method available, it is the one people keep doing. A weighed-and-recorded diary is more accurate for about ten days, which is roughly when most people abandon it.

Photograph before eating, from a slight angle rather than straight down, with something for scale in the frame — a fork or your hand will do. Include the whole plate rather than a close-up of the best-looking part.

Say or type what a photo cannot show: the cooking fat, the sauce, whether the yoghurt was full-fat. Scan the barcode when the food came out of a packet. And edit the result when you know better than the estimate — every CalEasy meal breaks into ingredients you can adjust in seconds, which is both faster than manual entry and more accurate than accepting an estimate you doubt.

For foods you eat constantly, it is worth getting the numbers right once and saving the meal. After that it is one tap, with no estimation involved at all.

Barcode scanning is the most accurate for packaged food and requires no judgement. Weighing ingredients and entering them manually is the most accurate for home cooking, and by far the slowest. Photo logging is the fastest and handles food you did not cook. Voice and text sit in between: quick, and able to express things a camera cannot see.

None of these is the right answer for every meal, which is why the useful question is not "which method is most accurate" but "which method will I actually use for this meal, right now". A rough estimate that gets logged beats a precise one that does not.