Snap your food. Know your calories.

Identification. The model names what is on the plate. This part works well: grilled chicken reads differently from poached, roasted vegetables carry visible oil, a croissant is unmistakable. These are dishes the model has seen thousands of times.

Portion estimation. The model judges how much there is. This is the hard step, and it is hard for a reason that has nothing to do with AI: a photograph is flat. Depth has to be inferred from shadow, from the rim of the plate, from the way the food sits. A heaped bowl of rice and a thin layer of it can look alike from directly above.

Nutrition lookup. Once the dish and the quantity are settled, calories and macros come from reference data rather than from a guess. That step is arithmetic.

So when people ask whether photo tracking is accurate, the honest answer is that one of the three steps is close to solved, one is arithmetic, and one is an estimate. The estimate is the portion.

We can be specific here, because we have our own history to look at rather than someone else's marketing.

That single number is the whole argument for photo logging. Four meals out of five are not one entry in a database; they are a protein, a starch, a vegetable and something poured over the top. In a search-and-select tracker that means four separate lookups, four portion guesses and four chances to give up. A photo handles it in one gesture.

About one analysis in seven gets edited by the person who sent it. We publish that figure deliberately, and we deliberately do not call it an error rate — because most of those edits are not corrections. They are additions: the oil that was used, the second slice, the sugar in the coffee, the fact that this was the large size. The camera could not have known.

It is also worth being clear about what precision is for. If your goal is a weekly energy deficit, what matters is whether your logged intake tracks your real intake consistently — not whether any single meal is exact to the gram. A method that is roughly right every day beats one that is precisely right on Tuesday and abandoned by Thursday. That is not an excuse for sloppiness; it is the reason adherence is the variable worth optimising.

Two habits make the numbers meaningfully better, and both take seconds:

• • Photograph before you start eating, from a slight angle rather than straight down. The angle is what gives the model depth.
• • When a dish is mixed, cooked in fat, or eaten out, add one line of text. "Two tablespoons of olive oil." "Large portion." That line is worth more than a better photo.

Most people who quit tracking do not quit because it stopped working. They quit because of the friction — the database search, the twenty near-identical entries, the kitchen scale that is at home when dinner is not. Accuracy that costs forty seconds a meal is accuracy you will abandon.

Our own usage says something about this. Breakfast is the most-logged meal at 31% of entries, lunch 25%, dinner 21%, with the rest spread across snacks. Breakfast wins because it is eaten at home, unhurried, from familiar ingredients. Dinner loses because that is when life interferes. Any method that only works under breakfast conditions will not survive a month.

So the thing to optimise is not precision. It is the number of seconds between finishing a meal and having it logged. Under ten, and tracking survives restaurants, travel, bad days and birthdays. Over thirty, and it doesn't.