To estimate calories from a picture, AI has to work out one thing above all: how much food is there. The same bowl of rice swings by 150 kcal depending on whether it is half full or heaped. Here is how to make it accurate.
A photo carries no depth information, so AI scales the food against a known reference in the frame. A Korean spoon is 19–20 cm long, chopsticks 22–24 cm, an adult palm 8–9 cm across. With one of those visible, bowl diameter and food height can be worked backwards and portions get far more accurate. Shoot the food alone and there is no way to tell one serving from two.
Straight down hides height; too low and the food at the back disappears. About 45 degrees, with the rim of the bowl in frame, works best. If there are several side dishes, get them all in one shot so nothing is missed.
If you skip the broth but the soup is counted in full, calories are badly inflated. The same goes for scraping off sauce, removing batter, or eating half the rice. Write it once in Food notes — for example "I don't drink the broth" — and every future analysis follows that rule.
If bibimbap was read as bibim noodles, just say so and it recalculates. Correcting the amount ("only half a bowl of rice") matters just as much as the name. Accurate records are what make the daily balance and the next-meal suggestion meaningful.
Photograph a whole shared spread and you get the calories of the entire table. Serving your portion onto your own plate first is the most accurate approach; if that is not practical, add a correction such as "shared between three people".
Weighing every meal to the gram lasts a few days at best. What matters is logging consistently. Even rough numbers, gathered daily, reveal the pattern behind the days you slip — and that is what moves the scale. CalorAI exists to make that habit as light as one photo.