Guide · Logging
Track calories with a photo: how accurate AI estimates really are
Snap your plate, read the calories. Three studies have compared those estimates with weighed meals, and the errors follow a clear pattern.

In short
Tracking calories with a photo is quick, but it gives you an estimate, not a measurement. In an NIH analysis of 102 weighed meals, four popular apps came in 250 to 345 kcal too low per meal on average, mostly because about 30 g of fat went missing. A photo is fine for a rough log. Add oil, sauces and big portions by text or barcode.
How accurate is calorie counting from a photo?
For a single meal, a photo estimate is often off by about a third, and usually on the low side. Three studies that compared photos with weighed food all point the same way.
The largest comes from the US National Institutes of Health. A team at NIDDK ran photos of 102 meals from a diet study through four widely used apps, including MyFitnessPal and Lose It!. The meals came from a metabolic kitchen where every ingredient was weighed to 0.1 g. All four apps undershot: on average by 250 to 345 kcal per meal and by around 30 g of fat. Carbohydrates were the most consistent. The results were presented at NUTRITION 2026 in July and have not yet been through peer review.
A group in Gothenburg tested the language models directly in 2025. ChatGPT-4o, Claude 3.5 Sonnet and Gemini 1.5 Pro each got 52 standardized photos in small, medium and large portions. For energy, ChatGPT and Claude were off by 35.8 percent on average, Gemini by 64.2 percent. The authors see the models as a screening tool, not yet good enough for precise dietary assessment.
The third study is from Dublin. ChatGPT-4 looked at 38 meals across 114 photos. It recognized what was on the plate well, with 93 percent precision. Portion weight was another story: it guessed too low on 76 percent of the photos.
So the AI usually knows what you're eating. How much of it is the weak spot.
Why does AI underestimate oil, sauces and big portions?
Because a photo only shows surfaces: a tablespoon of oil in the pan adds roughly 120 kcal and never shows up in the picture. Fat has 9 kcal per gram, more than twice as much as protein or carbs; the conversion is explained in weekly macros. The 30 g of fat missing per meal in the NIH data works out to about 270 kcal. That alone covers most of the gap.
Depth is the second problem. A photo is flat and a bowl of pasta isn't. In the Dublin study, large portions weighed 798 g on average (1.76 lb), while the estimate was 530 g (1.17 lb). Small portions were close: 408 g real, 431 g estimated. The Gothenburg team found the same thing. The bigger the plate, the bigger the undercount.
Then there's what nobody can see. Butter in mashed potatoes, cream in a sauce, cheese under the vegetables, sugar in the dressing. Stews, casseroles and smoothies are close to a lucky dip for any image model.
Sometimes the model simply gets the dish wrong. In Gothenburg, one model took scrambled eggs for pasta and another read falafel as meatballs. Those errors are rare, and you spot them right away if you glance at the ingredient list after taking the photo.
Example week: photo, chat and barcode combined
This week belongs to someone with a weekly budget of 15,400 kcal, or 2,200 a day. Hot meals get a photo, packaged food gets scanned, and anything with hidden fat or a drink gets a few typed words.
| Day | What they ate | Logged by | kcal |
|---|---|---|---|
| Mon | Oatmeal with milk, turkey sandwich, stir-fry with rice | barcode, chat, photo + "1 tbsp oil" | 1,850 |
| Tue | Greek yogurt, chicken burrito, cheese on toast | barcode, photo + "sour cream and cheese", chat | 2,150 |
| Wed | Eggs on toast, lentil soup, chicken with rice | chat | 1,700 |
| Thu | Bagel with cream cheese, pasta with pesto, salad with feta | barcode, photo + "2 tbsp pesto", photo | 1,900 |
| Fri | Ham sandwich, doner kebab, two pints | chat | 2,300 |
| Sat | Pancake brunch, burger with fries, two glasses of wine | photo + "large portion", photo, chat | 2,600 |
| Sun | Eggs and bacon, roast beef with gravy, an apple | photo, photo + "extra gravy", chat | 2,200 |
| Week | 8 photos, 21 meals | 14,700 |
At 14,700 kcal the week ends 700 kcal under budget. Saturday is the biggest day and Wednesday makes up part of it, which is exactly how a weekly calorie budget works.
Now the reality check. Say the person had accepted all eight photo estimates without adding anything, and each one was as far off as the NIH average of roughly 300 kcal. The app would show only 12,300 kcal, which looks like 3,100 kcal of room. The real number is 700.
Each correction takes a few seconds. A word like "fried" or "2 tbsp pesto" is enough for the estimate to include the fat the camera missed.
Two scenarios: desk job and building site
Emma, 65 kg (143 lb), office job. She eats lunch in the work cafeteria and dinner at home and photographs both. Her weekly budget is 12,600 kcal (1,800 a day) with a planned deficit of 2,800 kcal a week. Five of her ten weekday photo meals are heavy ones: creamy chicken pasta, mac and cheese, a breaded fillet. If each of those five is 250 kcal short, her real deficit shrinks to 1,550 kcal. Using the rough rule of 7,000 kcal per kilo, that's 0.9 kg (2 lb) in four weeks instead of 1.6 kg (3.5 lb). Why that rule is only approximate is covered in the 7,000 calorie deficit per week.
Jake, 90 kg (198 lb), construction worker. He burns a lot and eats big plates. His budget is 19,600 kcal a week (2,800 a day) with a planned deficit of 3,500 kcal. Dinner is often around 1,200 kcal. If the AI underestimates the weight by a third, as in the Dublin study, the app shows about 800 kcal. Over seven dinners that's 2,800 kcal missing each week, and only 700 kcal of the deficit is left. After a month, Jake can't see why the scale has barely moved.
For Jake, a kitchen scale pays off once. If his usual pasta plate weighs 650 g (23 oz), he types "650 g spaghetti bolognese" from then on instead of only taking a photo. For Emma, adding one word for the sauce or the breading is enough.
The week view shows what a gap like that does. The screenshot shows a deficit of 2,968 kcal, about 0.39 kg (0.86 lb) of fat. If five photo meals in that same week were each 250 kcal short, only 1,718 kcal would remain, just under 0.25 kg (0.55 lb).
Your weight trend tells you in the end whether your log is roughly right. If it drops much more slowly over three to four weeks than your deficit suggests, calories are probably missing from your log. Photo meals with sauce are the first place to look.

When are chat or barcode more accurate than a photo?
For anything with a nutrition label, the barcode wins, because the values come from the package and only the amount is left to guess. Think yogurt, cereal, protein bars, frozen pizza and bottled drinks.
Typing wins as soon as you know amounts or hidden ingredients. "200 g cooked pasta with 2 tbsp pesto" is easier for any model than a photo of the bowl, because the grams and the fat are already there. Drinks are better as text too: a pint of lager and a pint of cider look much alike in a picture.
A photo works well for plates with clearly separate parts, such as meat, a side and vegetables, side by side and easy to see. It's also handy at restaurants, where you don't know the recipe anyway. There, an estimate you then correct beats no entry at all.
Rule of thumb: the more fat is hidden in the food, the more a short line of text is worth.
Common mistakes and honest limits
- Accepting the result without a look. Read the ingredient list after each photo. If oil, butter or sauce is missing, add it.
- Forgetting drinks. Beer, wine, lattes and juice rarely make it into the plate photo. Log them by chat.
- Photographing only the first plate. Seconds, the bread basket and dessert count too, or a Sunday roast can end up logged at half its size.
- Not flagging big portions. The undercount grows with plate size. Writing "large portion" or a gram amount makes up for part of it.
- Panicking over every error. Single mistakes partly cancel out. What matters is the weekly total and the trend over several weeks, see weekly calorie balance.
Now the limits. The studies are small, the NIH data is still preliminary, and each tested specific apps and model versions. Every model estimates differently, and there is no published accuracy test of the AI inside Weekfuel. Treat any number from a photo as an estimate.
Where precise values matter, a photo isn't enough. That applies to insulin dosing with diabetes, to pregnancy, or if food and counting feel stressful for you. In those cases, plan with your doctor or a registered dietitian. And however accurate your log is, don't plan days below 1,200 kcal (women) or 1,500 kcal (men) without medical supervision.
How to log by photo, chat and barcode in Weekfuel
In Weekfuel you photograph your plate or pick an image from your library. The AI recognizes the parts, estimates amounts and splits the meal into ingredients, just like with a typed message. The screenshot shows the result for "chicken with rice": 225 g chicken breast, 90 g rice, 353 kcal and 55 g protein.
If something is missing, just send a follow-up such as "fried in 1 tbsp olive oil" or "it was a large portion". You can also adjust amounts with plus and minus. For packaged food you scan the barcode, and the values come from Open Food Facts. The app doesn't keep the photo. More on the chat calorie tracking page.
Every entry counts toward your weekly budget, and steps from Apple Health come in as calories burned. Each day you can see how much of the week is left. Weekfuel is iPhone only and needs the Pro annual subscription; there is no free version.

Frequently asked questions
How accurate are calorie apps that use photos?
In a preliminary NIH analysis of 102 weighed meals, four popular apps were 250 to 345 kcal too low per meal on average. They usually recognize the dish, but they are much weaker on amounts and fat.
Can ChatGPT count calories from a photo?
It can estimate them. In a Gothenburg study, ChatGPT-4o and Claude 3.5 Sonnet were off by about 36 percent on energy on average, with larger errors for bigger portions.
Why does my app show fewer calories than I actually ate?
Usually hidden fat like oil, butter or cream is missing, or the portion was bigger than it looks in the photo. Add those ingredients and "large portion" in text.
Is photo tracking more accurate than a classic food diary?
About the same. The Gothenburg researchers found ChatGPT and Claude close to traditional self-reporting. It gets more accurate when you add amounts and hidden ingredients to the photo estimate.
Do I still need a kitchen scale?
Not for every meal. It does help to weigh your usual portions a few times, such as pasta, rice or cereal. After that, type the grams instead of letting the AI guess.
Does Weekfuel store my food photos?
No. The photo is sent to the AI service for analysis and isn't kept by the app. The meal it recognized then stays on your iPhone only.