How does photo calorie tracking work?
Photo calorie tracking uses a computer-vision model to identify the ingredients of a meal from a single photo of the plate, then estimate calories and macros without weighing or manual entry. It is an alternative to a traditional food diary that multiplies long-term adherence by 3 to 5.
How it works technically
The classic pipeline combines several models. A first detection model isolates the plate in the image and segments the different zones (starches, protein, vegetables, sauces). A second model, trained on hundreds of thousands of dish images, identifies each food category. A third model estimates the volume of each portion from perspective and the relative size of the plate.
The detected volumes are then matched against a nutritional database (Ciqual-type, USDA, or proprietary databases) to produce calories, macronutrients, and micronutrients. The result appears in a few seconds, ready for the user to confirm or adjust.
Current accuracy, in 2026, is ±15 to 20% on common, well-lit dishes, and it drops for hidden mixed dishes (gratins, sauced dishes, soups). It is less precise than weighing gram by gram, but more than enough for daily tracking and clearly more reliable than a food diary filled from memory in the evening.
Why it is useful
The main added value is adherence. Studies of manual food diaries show a high dropout rate: more than half of users stop within two weeks, mainly because of the friction of repeated logging. A photo takes three seconds; logging a food in MyFitnessPal takes thirty to ninety.
Photo tracking is also more honest. Average calorie underreporting in manual diaries is 20 to 30% according to the literature. A vision model does not spontaneously shrink portions, which gives a more realistic picture of actual intake.
Finally, it is the only method suited to unplanned meals: a restaurant, an improvised dish, brunch at a friend's. There is nothing to weigh or break into components — one shot is enough.
Limits to know
Accuracy is not that of a scale. For a high-level sports goal or a supervised therapeutic diet (kidney failure, type 1 diabetes), a 15–20% margin of error can be too wide. For most goals (loss, cut, maintenance, muscle gain, everyday healthy eating), it is more than enough.
Hidden mixed dishes are the worst case: a cassoulet under its crust, a parmigiana, a lasagna — the model can only detect what it sees. Manually entering the dish name can then complete the photo.
Privacy deserves a check: some apps send photos to remote servers for processing. Look at the privacy policy, especially if the photo captures a personal setting.
How Daizu does it
In Daizu, meal scan is included in the subscription, at 200 scans per day. The user takes a photo, the result appears in a few seconds with calories and macros, and it is added automatically to the day's total aligned with the goal.
Images are not kept beyond processing and are not used to train advertising models or for profiling. Policy details are on the app's privacy page.
The scan complements the program: for dishes Daizu generated, calories come straight from the recipe (no photo needed); the scan is only for deviations and unplanned meals.
