Honest numbers

Every calorie number you have ever logged was an estimate. This one admits it.

A photograph is not an accurate way to count calories. Neither is a person looking at the same photograph, or a database entry for a food that varies, or a weighed log that leaves out the frying oil. Snapkin shows you which part of the number it is confident about and which part it made up.

€14.99 a month, or €99.99 a year. Three days free to start, no free tier, no upsells later. The App Store or Google Play takes the payment and shows you the price and tax for your country before you confirm.

A logged meal marked "Best guess": 530 kcal, the four macros, and two ingredients with their grams and calories, one of them noted as seared in about 8 g of oil folded into the line, above a box naming the frying fat and the buttering of the mash as the biggest unknowns.
Every line shows its working
A first meal result: chicken with rice and broccoli at 580 kcal marked "Solid estimate", the macros, and the four ingredients it found with the grams and calories it gave each, including the butter it was fried in.
What comes back, first time

The published numbers, including the ones we would rather not print

These are not our figures. They are what the literature reports for vision models estimating meals from photographs, and the page they are argued out on is linked at the bottom.

24–34% Energy error per meal ACETADA, 806 meal images against food weighed to 0.1 g. Roughly 166 to 211 kcal.
35.8% Energy error, second study Fridolfsson et al., 52 photographs against a calibrated scale.
41% A nutritionist, same photos From the Nutrition5k study. Non-nutritionists were 53% out.

A photograph is not accurate. Neither is a person looking at the same photograph. Nobody is choosing between this and a perfect log. They are choosing between this and no log at all. What the research does show is where the error lives: recognising the food already runs at 74 to 93%, and portion size is the bottleneck. So the app asks you a question back instead of handing over a total, and writes down what it assumed.

Read the full argument, with the evidence against

Honest numbers

A photo gives a good guess, not a lab result. Snapkin says which is which.

Every ingredient shows the assumption behind it, and the meal says how sure it is overall. When it is guessing, it says so. When it is wrong, you say so, and it shows the before and after.

The meal carries a confidence.

Best guess, or mostly assumed -- a summary of how much of the plate it could see and how much of it it had to work out.

Each ingredient carries its assumption.

"A whole fillet, sized against the plate." "Seared in about 8 g of oil, folded into this line." You can see the sentence that produced the number.

It names its biggest unknown out loud.

Usually the cooking fat, sometimes the portion. That is also the question it will ask you, and answering it is the highest-value thirty seconds you can spend on a meal.

Correcting it shows the before and the after.

Say the portion was smaller, or that it was fried in butter, and the meal redraws with the change marked. Every correction is undoable and every one of them is also something it learns.

See a meal itemised

That is the whole habit, and it survives restaurants, holidays and Thursdays, because there is nothing to keep up.

€14.99 a month or €99.99 a year, billed by the store, cancel there any time.

What it costs