Every calorie count is an estimate. This one admits it.
- Tested on 500 meals weighed on a scale.
- Within 17% on a typical meal.
- It shows what it assumed, so you can fix it.
No card for the first 3 days. €9.99 a month or €99.99 €49.99 a year, via the app.launch price, for a limited time
The photo gets you close. The question covers what the camera cannot see.
Point the camera. A few seconds later there is a number. Sometimes one thing the camera cannot see would change it, like butter in the pan. Then Snapkin asks you. Answering takes one tap.
It does not hide what it assumed.
A camera cannot see the oil in the pan, or the weight hiding under a sauce. Snapkin does not quietly pick a number and show it as fact. It tells you what it assumed.
It only asks when the answer changes the calories.
At most three questions, and often none. It only asks when the answer would change the calories by a real amount.
Answering is a tap, not a form.
A few likely answers and "I do not know". Type your own if none fit. No searching a database, no weighing anything. You can skip it and the meal still logs.
An example meal. What we measured on 500 weighed meals is in the next section.
We checked Snapkin against 500 meals weighed on a scale
Researchers weighed and photographed each meal. We sent each photo to Snapkin like the app does, and compared the calories with the scale. No questions answered: this is the first number you see.
We ran this test ourselves, so keep that in mind. Not peer reviewed. The meals are from three public research sets. We run all 500 again before changing the AI model, and publish the result.
Your scale corrects what is left. Counts can be a bit high or a bit low. It depends on your meals. The monthly check-in measures your target in the same counts as your meals. So the error cancels out, whichever way it goes.
Close enough to track with. And it shows you what it assumed.
Every ingredient shows what Snapkin assumed about it. The meal says how sure it is overall. When it is guessing, it says so. When it is wrong, you say so. Then it shows the before and the after.
The meal shows how confident it is.
It says "Solid estimate", "Best guess" or "Rough guess". That sums up how much of the plate it could see, and how much it had to work out.
Each ingredient shows its assumption.
"A whole fillet. Size worked out from the plate." "Fried in about 8 g of oil, included in this line." You can see the sentence behind the number.
It names its biggest unknown.
Usually the cooking fat, sometimes the portion. If it could change the calories a lot, it asks you about it. Answering takes one tap.
Correcting it shows the before and the after.
Say the portion was smaller. Or say it was fried in butter. The meal updates, and the change is marked. You can undo it. Overnight, it may keep what you told it as a sentence in its memory.
No card for the first 3 days. €9.99 a month or €99.99 €49.99 a year, via the app.launch price, for a limited time
And when it is wrong, you say so
You do not edit a number in a form until it agrees with you. There is a box on every meal. You type what you would have said out loud.
- The rice was more like half that.
- There was a lot more oil in the pan than that.
- That was the light mayonnaise, not the normal one.
- The scale was weighing the bowl as well.
- I only ate about two thirds of it.
- It was a deep bowl, not a plate.
A before and an after, per ingredient.
It does not just say "the meal is now 858 kcal". It says "roast potatoes, 260 → 200 g", with the new total beside it. You can see which of your words did what.
It says what it had assumed.
Its reply often says what it had guessed before. The old numbers stay in the history. So you find out what it had been guessing about your kitchen.
It is kept on the meal.
"What changed · 1 correction · +82 kcal" sits under the ingredients. Open it and the whole history is there.
Nothing you tap is final.
Undo the last correction, and the meal goes back to what it was. You can undo every answer and every correction, one at a time, newest first.
What the published studies measured, and when they measured it
These are not our numbers. They are what published studies found for other software that estimates meals from photos. The page that argues with them is linked at the bottom.
These studies tested older AI models. ACETADA and Fridolfsson tested models from 2024 and 2025. On phone photos like ACETADA's, Snapkin was 19% out on average. The tests differ, so it is not a head-to-head.
A photo does better than a person guessing. Nutritionists looking at the same photos were 41% out on weight. A photo cannot show everything, like the oil in the pan. So check what Snapkin assumed.
See a meal itemised
It's that simple. It keeps working through restaurants, holidays and busy weeks. There is nothing else to keep up with.
No card for the first 3 days. €9.99 a month or €99.99 €49.99 a year, via the app.launch price, for a limited time