How to Count Calories From a Photo
To count calories from a photo, you take a picture of your meal and an AI model identifies each food, estimates the portion, and matches it to nutrition data to return calories and macros in seconds. It is an estimate, not a measurement: studies put average errors between under 1% and about 38%, depending on how complex the plate is.
That range sounds wide, and it is. But the alternative most people actually use is not a lab-grade food scale. It is a guess typed into a database, or no log at all. This guide explains how photo calorie counting works, where it is reliable and where it is not, and how to get the most honest numbers out of it on the food you really eat, from a nasi lemak bungkus to a plate of chicken rice.
Why do people quit manual calorie tracking?
Manual logging has a completion problem. Every meal asks you to do three things well: name the food correctly, pick the right database entry, and estimate the portion. Miss one and the number is wrong. Do all three, several times a day, for weeks, and the effort adds up until you stop.
It is worse with Southeast Asian food. Type "nasi campur" into a typical database and you get dozens of entries that have little to do with your plate, because nasi campur is whatever you scooped from twenty trays that day. The friction is not a discipline problem. It is a design problem.
Photo logging collapses those three steps into one thing you already do without thinking: pointing your phone at your plate. The best tracking method is not the most precise one; it is the one you will still be using in a month.
How does AI food recognition work?
A photo calorie counter does two jobs at once: working out what the food is, and estimating how much of it there is.
- Recognition. The image runs through a computer-vision model trained on large numbers of labelled food photos. It learns the visual signatures of dishes, like the glossy dark noodles of char kuey teow or the separate mounds of rice, sambal and egg in nasi lemak, and returns its best guess for each item on the plate.
- Portion estimation. From a flat 2D photo, the app infers volume using cues such as plate size, pile height and familiar objects in frame. This is the harder step, and it is where most of the error comes from.
- Nutrition lookup. Each recognised food is matched to nutrition data and multiplied by the estimated portion to give calories, protein, carbohydrates and fat.
Newer apps add a fourth step: you can add a sentence of context ("extra sambal", "no skin", "half the rice") and the estimate updates. That small correction often matters more than the photo itself.
How accurate is photo calorie counting?
Honest answer: useful, not exact. A 2023 systematic review in Annals of Medicine pooled 52 studies of AI image-based dietary assessment. Average calorie errors ranged from 0.10% to 38.3% across studies, and the authors found errors were lower when images showed single, simple foods. Mixed dishes such as curries, and plates with overlapping foods, carried a higher risk of misclassification. Overall, AI performance fell within the range of trained human estimators.
In practice, that means:
- A banana, a boiled egg or a slice of bread will usually be estimated well.
- A plate of rice with two clearly separated dishes is a fair test, and the result is usually close enough to plan with.
- A bowl of curry mee or a heavily sauced nasi kandar is the hardest case, because the gravy hides the protein and the oil.
The right mental model is a knowledgeable friend glancing at your plate and giving a quick, sensible estimate. Over days and weeks, random over- and under-estimates partly cancel out, and the trend is what drives your results. Consistent errors, such as always missing the cooking oil, do not cancel. That is why the tips below focus on those.
Where does photo calorie counting go wrong?
Four situations account for most bad estimates:
- Hidden fat. Oil, santan, butter and ghee cooked into a dish are invisible. One tablespoon of palm oil is about 120 kcal according to USDA FoodData Central, and a wok-fried hawker dish can easily carry more than one.
- Hidden sugar. The condensed milk in teh tarik or the gula melaka in a dessert does not show up in a picture.
- Stacked or deep food. A tall burger or a deep bowl of noodles hides what is underneath, so volume is hard to judge.
- No sense of scale. Food filling the whole frame, with no plate edge or cutlery, gives the model nothing to anchor portion size against.
How many calories are in common Malaysian dishes?
Knowing a rough reference value helps you spot an estimate that is clearly off. These figures come from published sources and are for the specific serving sizes listed. Your stall's portion may be bigger or smaller.
| Dish | Serving analysed | Approx. calories | Source |
|---|---|---|---|
| Nasi lemak (with sides) | 395 g plate | ~740 kcal | Yeo et al., Foods 2021 |
| Steamed chicken rice | 412 g plate | ~780 kcal | Yeo et al. 2021 |
| Fried kuey teow | 380 g plate | 670 kcal | Malaysian Dietary Guidelines 2020 |
| Fried mee | 300 g plate | 519 kcal | Malaysian Dietary Guidelines 2020 |
| Roti canai | 1 piece (84 g) | 266 kcal | Malaysian Dietary Guidelines 2020 |
| Teh tarik | 1 mug (250 ml) | 140 kcal | Malaysian Dietary Guidelines 2020 |
The coconut rice in nasi lemak alone is about 169 kcal per 100 g in the Malaysian Food Composition Database (MyFCD), so the rice is only part of the story; the fried chicken, sambal and peanuts do the rest. For a dish-by-dish breakdown, see our guides to nasi lemak calories and mamak food calories.
If your app says a full plate of char kuey teow is 350 kcal, that is a sign to check the portion or add the oil.
How do you get the most accurate photo calorie estimate?
These habits take no extra time and fix most of the errors above:
- Shoot at a 45-degree angle. A three-quarter view shows both surface area and height, which helps portion estimation. Straight-down shots flatten a mound of rice.
- Keep a size reference in frame. A spoon, fork, standard plate or your hand gives the model something to scale against.
- Use even light. Daylight beats a dim stall under yellow bulbs; heavy shadows and colour casts confuse recognition.
- Photograph before the first bite. Capture the full portion, not the half-eaten aftermath.
- Add one line of context. "Fried, not steamed", "kurang manis", "extra rice" or "two pieces of chicken" covers what the camera cannot see.
- Split mixed plates. For nasi campur, check that each item is listed separately so the fried fish, the curry and the vegetables are each estimated on their own. If you eat economy rice often, our guide to losing weight on hawker economy rice shows which trays add up fastest.
Photo vs manual entry vs barcode: which is best?
None is best everywhere; they fit different situations.
| Method | Best for | Speed | Effort | Accuracy notes |
|---|---|---|---|---|
| Photo (AI) | Hawker, restaurant, home-cooked and mixed plates | Seconds | Very low | Good on visible food; misses hidden oil and sugar unless you add context |
| Manual entry | Simple foods and recipes you repeat | Slow | High | Only as good as the database entry you pick and your portion guess |
| Barcode scan | Packaged, labelled products | Fast | Low | Precise for the labelled serving; useless for unlabelled food |
| Food scale | Competition prep, clinical diets | Slowest | Highest | Most exact for raw ingredients; impractical when eating out |
Most people in Malaysia and Singapore eat out several times a week, often at stalls with no nutrition labels. That is where a photo, or a quick text or voice description, is often the only practical option. Packaged snacks are where a label or barcode does better.
How do you start logging meals from photos?
You do not need a new routine, only one small swap. At your next meal:
- Before eating, take one clear photo at an angle, with a spoon or the plate edge visible.
- Let the app identify the items and read the estimate.
- Fix anything obviously wrong and add a line about cooking method or extras.
- Save it and move on. The whole loop takes less time than searching a database.
Then look at your weekly average, not individual meals. If your daily calorie target still needs working out, the TDEE calculator and our guide on how many calories to eat to lose weight cover that. Once calories feel easy, tracking macros is the natural next step.
If you want to try this on your own meals, Berry Best lets you log food to Berry, its AI coach, by photo, text or voice, then shows calories and macros you can correct in the same chat.
Sources
- Shonkoff E. et al. (2023). AI-based digital image dietary assessment methods compared to humans and ground truth: a systematic review. Annals of Medicine.
- Yeo P.L.Q. et al. (2021). Energy content and nutrient profiles of frequently consumed meals in Singapore. Foods.
- Ministry of Health Malaysia. Malaysian Dietary Guidelines 2020, Table 2.4.
- Malaysian Food Composition Database. Rice, coconut milk (nasi lemak).
- USDA FoodData Central. Oil, palm.
Keep reading
- How many calories should you eat to lose weight?: set the target your photo logs are measured against.
- 7 calorie-counting mistakes that stall fat loss: the hidden oil and drinks that photos miss.
- Nasi lemak calories: what each component of the plate adds.
- Hawker economy rice for weight loss: building a lighter nasi campur plate.
- Berry Best vs MyFitnessPal: photo-first logging compared with a database-first tracker.
- Best calorie-tracking apps compared: an honest round-up, Berry Best included.
Frequently asked questions
- Can you count calories from a picture?
- Yes. A photo calorie counter uses image recognition to identify each food on the plate, estimates the portion from visual cues such as plate size and pile height, then looks up nutrition data to return calories and macros. The result is an estimate, so you should glance at it and correct anything the app misread, like the wrong dish or an oversized portion.
- How accurate are AI calorie counting apps?
- Accuracy depends heavily on the meal. A 2023 systematic review in Annals of Medicine found average calorie errors ranging from under 1% to about 38% across studies, with simple single foods doing far better than mixed dishes like curries. That puts AI roughly in the same range as trained human estimators: good enough to guide daily choices, not precise enough for clinical use.
- Can AI apps recognise Malaysian food like nasi lemak?
- Most modern food-recognition models handle common Malaysian dishes such as nasi lemak, roti canai and chicken rice, but local plates are a harder test because sambal, gravy and oil hide what is underneath. Expect the dish name to be right more often than the portion. Adding a quick note, such as 'extra sambal, fried chicken thigh', gives the estimate a lot more to work with.
- Is photo logging better than weighing food?
- Not for precision, but usually for consistency. A kitchen scale is the most exact way to measure food and is worth using for competition prep or medical diets. For everyday weight loss, photo logging wins because it takes seconds and works at a hawker stall where a scale is impossible, and a reasonable log you keep every day beats a perfect one you abandon.
- Why does my food photo calorie estimate seem too low?
- The usual reason is hidden fat. Cooking oil, santan, butter and sugar dissolve into the food and are invisible in a photo, so the estimate covers what the camera sees rather than everything in the dish. Fried, gravy-heavy and stir-fried meals are most affected. Tell the app how the food was cooked, or add the oil as its own item, to close the gap.