Calories from a photo: how the AI works and how far you can trust it
The biggest enemy of calorie counting isn't lack of knowledge — it's friction: weighing, searching databases, tapping. After two weeks, most people give up. Photographing your meal cuts that friction to almost zero — as long as you know how it works, where it goes wrong, and how to help it.
How AI counts calories from a single photo
Modern vision models do three things at once:
1. They recognize what's on the plate. Not "food" but specifics: chicken breast, rice, carrot salad, yogurt sauce. The model also sees how it was prepared — breaded vs grilled is a completely different calorie count.
2. They estimate portion size. From proportions on the plate, against cutlery, a hand or standard dishes. This is the hardest part and the main source of error — which is why there's a whole section below on how to take the photos.
3. They map it to a nutrition database. The recognized ingredients and weights become calories, protein, fat and carbohydrates — item by item, not "in one shot", so you can correct a single ingredient instead of the whole meal.
How accurate is it? Honestly
The typical error for a single meal estimated from a photo is ±20–30%. Sounds weak? Context is everything: doubly-labeled-water studies have shown for decades that people underreport by hand — their intake by 20–40%, systematically and unconsciously (underestimated portions, "forgotten" snacks, oils, sauces). The AI's error is of a similar size, but random, not directional — sometimes high, sometimes low — and it has no motivation to flatter you.
What matters more, though, is something else: in food logging, what counts is consistency, not the perfection of any single entry. If your counter is wrong in a reasonably stable way, then your real energy needs computed adaptively from your weight trend automatically "absorb" that error: the TDEE↔target math is calibrated on the same data, carrying the same error. A consistent log with a ±25% error beats an on-and-off log of perfectly weighed meals — every time.
How to take photos so the AI counts better
Frame the whole meal at a slight angle (~45°). A perfect top-down view flattens depth (bowls!), a side view hides the contents. The angle shows both the surface area and the height of the portion.
Leave something for scale in the frame. A fork, a hand, a standard plate — anything of known size. It's the single biggest accuracy improvement for portion estimates.
Light and focus. A photo taken in a dim restaurant will still be counted, just with more uncertainty. The second it takes to lock focus pays off.
Drinks and second helpings separately. A milky latte, juice, a second scoop of potatoes — those are the calories that vanish most easily. Add them as a separate entry or mention them in the description.
Describe "layered" dishes in words. Lasagna, casserole, a thick soup — the AI can't see what's inside. One sentence ("with meat, lots of cheese") can correct the estimate by tens of percent. In Bilberry you can combine photo and text in one entry.
When to correct manually (a short list)
A photo can't see four things — and they're exactly what most often "disappears" from food logs:
1. Cooking fat — a tablespoon of oil is ~90 kcal, and you can't see it on the pan.
2. Dressings and sauces added afterwards — especially on "healthy" salads.
3. Sugar in drinks — two teaspoons in coffee ×3 a day = ~120 kcal.
4. Alcohol — 7 kcal/g, the second most calorie-dense macronutrient.
The good news: these are corrections you make once — after that, favorites memory takes over (repeat meals in one tap) plus a barcode scanner for packaged products.
Privacy: what happens to the photo
It's worth asking every app this question. In Bilberry, the photo is used to estimate the meal and is stored in your private log; you always confirm the computed entry yourself — nothing lands in your log "on its own". Health data is processed in line with the GDPR (details in the privacy policy) and is never sold to anyone — the business model is a subscription, not ads.
Frequently asked questions
Can the AI handle home cooking? Yes — chicken with rice, a stew or porridge is everyday work for vision models. The more "hidden" the structure of a dish (casseroles, thick sauces), the more a one-sentence description helps.
Do I have to weigh my food? You don't — for a typical weight-loss goal, photo accuracy plus your weight trend is entirely enough. A kitchen scale makes sense selectively: for oil, nuts and nut butter, where calorie density is extreme.
What about restaurant food? Photograph as usual and assume that restaurant versions are more caloric than homemade ones (more fat) — when in doubt, round up or add "cooked in butter" to the description.
What if the AI gets it wrong? You edit every entry before saving — correcting an ingredient, a weight, or the whole thing. The entry waits for your confirmation, so an error never "leaks" into your statistics without your knowledge.
Important: this article is for information only, and AI calorie estimates are approximate. They do not replace advice from a doctor or dietitian and are not for making medical decisions.
Snap. Confirm. Done.
In Bilberry you log a meal with a photo in seconds — AI recognizes the ingredients and calories, you just confirm. Plus an adaptive target and a weekly meal plan. 14 days free, no card required.
Try Bilberry →