How accurate is AI calorie counting from a photo?

Short answer. Counting calories from a photo works like this: the model identifies what is on the plate, estimates the size of the portion from visual cues, and matches it against nutrition data, so you get calories and macros back within a few seconds. It is an estimate, not a lab analysis. For ordinary meals the error is usually 10–30% per entry, and it drops noticeably if you add one line about the portion, the oil and the sauce. For everyday tracking that is plenty. For laboratory precision you need a scale.

What happens after you send the photo

1. Recognition. A multimodal model — Fat Track uses OpenAI models — looks at the image and names the components: not "food", but "rice, a chicken patty, cucumber, sour cream".
2. Portion estimate. From the size of the plate, the cutlery in frame, the height of the pile and the density of the food, it guesses the weight of each component. This is the hardest part and the one that goes wrong most often.
3. The maths. Calories, protein, fat and carbs per 100 g are pulled for each component, multiplied by the estimated weight and added up.
4. Entry and correction. The result is saved to your diary. Reply "portion 300 g" or "fried in oil" and the numbers are recalculated.

A typed or spoken description goes through the same steps minus the first one: the model works straight from the names and weights you gave it.

How accurate it really is

Worth being straight about this. No system, ours included, can see how much oil soaked into a patty or how much sugar went into the lemonade. Published work on food-recognition systems and our own logs point to roughly this picture:

For comparison: in self-report studies, people keeping a diary by eye without scales underreport what they ate by 20–30% on average. A careful photo entry with corrections is about as accurate as a manual diary without a scale, and it takes seconds instead of minutes — which is precisely why it survives as a daily habit.

Where the model goes wrong most often

• Oil and frying. A tablespoon of oil is 135–150 kcal and it is invisible in the photo. Write "fried in oil" or "no oil".
• Sauces and dressings. Mayonnaise, creamy and cheese-based sauces can double the calories in a salad.
• The depth of the bowl. A soup in a deep bowl and a shallow one look identical from above. Shooting at an angle helps.
• Sugar in drinks. Coffee with syrup and coffee without are the same picture. Describe drinks in words.
• Stacked portions. Pancakes, fritters, slices of pizza: say how many. "Three pancakes" removes the guesswork.

Getting the most out of a photo

1. Shoot from above or at about 45 degrees, in normal light, with the whole plate in frame.
2. Leave something familiar for scale next to it — a fork, a spoon, your hand.
3. Add what the camera cannot see: cooking method, oil, sauce, number of pieces.
4. For home meals you eat often, do the maths properly once — weigh it — then just state the portion in text from then on.
5. Do not chase perfection. A steady 10–15% error does not stop you losing weight, as long as you log everything and correct your target by the trend on the scale.

When a photo is enough and when you need a scale

A photo is enough if the goal is to hold a deficit or simply to see what you are eating: errors in both directions partly cancel out across a week, and your weight trend tells you where to move the target. Scales and precise entries earn their place when you are preparing for a competition, following a medical diet, or stuck on a plateau and hunting for the leak. The practical answer is to combine them: photos and text every day, a scale for the home meals that repeat.

Judging portions with your hands and your kitchenware, for the days when there is no scale at all, is covered in how to count calories without a scale.

What this looks like in Fat Track

In the Fat Track bot for Telegram and MAX you send a photo, a line of text or a voice note, get calories and macros back, and can correct the entry in a reply. Checked values for basic foods, with fibre and glycaemic index, sit in the food database, and the history and charts live in the app. The full walkthrough of the scenario is on the calories from a photo page.

Common questions

Can I photograph a label instead? Better to type the product name and the grams — the estimate comes out closer than reading the packet from an image.

Does it handle home cooking — stews, casseroles, curries? Common dishes are recognised. For layered salads and thick soups, spell out the main ingredients.

What if the estimate looks too high? Reply with a correction: portion weight, cooking method. If you keep doubting the same dish, weigh it once and remember the reference.

Try it at your next meal: photograph the plate, send it, and compare the estimate with what you expected. The first seven days are free.