This is the first question anyone sensible asks, and it deserves a straight answer rather than a marketing one.
A photograph-based calorie estimate is genuinely useful and genuinely imprecise, and the two facts do not cancel out. What matters is understanding which parts of the estimate are solid, which parts are guesswork, and whether the error is small enough for the decision you are making.
What a model reads well
Identification is the strong part. Given a clear photo of a plate, a vision model is good at naming what is on it and separating it into components — this is rice, that is grilled chicken, there is a dressed salad next to it. That is a genuinely hard task that a food database search does badly, and it is where most of the time saving comes from.
Relative portion size is decent too, especially with a reference object in frame. A plate, a fork or a hand gives scale, and a dish photographed at a normal angle from a normal distance reads within a sensible range.
What it cannot see
Three things, and between them they account for most of the error.
Fat that has already been absorbed. Oil in a pan disappears into what is cooked in it. A tablespoon is around 120 kcal and it is completely invisible in a photograph. The same fried rice can differ by several hundred calories depending on a cook you cannot see.
Density and what is underneath. A bowl shows a surface, not a volume. Whether that surface sits on two centimetres of rice or six is a guess informed by the shape of the bowl, and the guess is sometimes wrong by a lot.
Hidden ingredients. Sugar in a sauce, butter finishing a steak, cream in a soup. None of it is visible, and none of it is small.
Where that leaves you
The practical picture is roughly this: a photo estimate is usually closer than eyeballing, usually further off than weighing, and its error is largest exactly where cooking fat is largest — restaurant food, fried food, anything glossy.
That is not a small caveat. But it is worth comparing against the real alternative, which for most people is not a kitchen scale. It is not logging at all, or logging from memory at the end of the day, which is reliably the least accurate method there is.
Getting more out of it
A few habits close most of the gap, and none of them cost much:
- Shoot from a slight angle rather than straight down, so depth is visible.
- Keep something of known size in frame — a fork, a standard plate.
- Say or type what the photo cannot show. "Fried in about a tablespoon of oil" is a one-second correction worth more than any amount of model improvement.
- Weigh the handful of things you eat constantly. Learning what 40 g of oats or 150 g of chicken looks like once pays out every day after.
- Correct the estimate when you know better. It is a starting point on a form, not a verdict.
The honest summary
If you need laboratory precision, weigh your food. If you need a number good enough to run a deficit against and correct from the scale over three weeks, a photo estimate does that job, and it does it at a cost low enough that you will still be doing it in March.
Consistency is worth more than precision here, and it is not close. What matters is that the same method is applied every day, so that when your weekly average stops moving you can trust the trend even if you cannot trust any individual number.