“Snap a photo, get your calories” is the promise of a whole category of apps. It’s genuinely faster than typing every ingredient, but a photo can’t weigh food or see the oil it was cooked in. This page collects what independent tests have found, what each popular photo app lets you check and fix, and which accuracy claims come with a method someone else could check.
We make Calorix, one of the apps in the table, so we’re not neutral. We make no accuracy claim for Calorix here, every fact about other apps links to its source, and everything was checked on October 11, 2026.
What independent tests found
The 2026 NIH test of four popular apps
Researchers at the US National Institute of Diabetes and Digestive and Kidney Diseases ran standardized photos of 102 meals through the photo features of four apps. The meals came from a controlled feeding study, where every ingredient is weighed to the nearest 0.1 gram, and averaged 918 kcal (abstract, ASN release).
| App (as tested) | Average under-count per meal | 95% confidence interval | Range for single meals (limits of agreement) |
|---|---|---|---|
| Appediet | 252 kcal | 210 to 295 | 672 under to 167 over |
| MyFitnessPal | 327 kcal | 269 to 385 | 884 under to 246 over |
| Lose It! | 333 kcal | 282 to 383 | 835 under to 170 over |
| Cal AI | 345 kcal | 296 to 392 | 821 under to 132 over |
All four apps under-counted by about a third on average, and fat by about 30 g per meal; carbohydrates were the most consistent. Early results from more than 200 additional meals suggested the apps struggle more with high-fat ketogenic meals. Two cautions: this is a conference abstract, so it’s preliminary until a peer-reviewed paper appears, and the app versions and test dates weren’t reported. One of the researchers, Aaron Hengist, advised that people using a photo-based app “without adjusting the portions or entering the amounts of food should take the results with a grain of salt.”
Earlier studies
- A 2024 study of seven apps on 22 studio photos found MyFitnessPal identified 97% of food components, but that’s recognition, not calories: on energy, MyFitnessPal under-counted by 3% and Foodvisor by 47% on average, and a plate of eggs on toast with butter was under-counted by 35% and 73%, because the butter was missed.
- General AI chatbots do no better: given 52 standardized food photos, ChatGPT-4o and Claude 3.5 Sonnet missed calories by 35.8% on average, Gemini 1.5 Pro by 64.2%, and all three under-counted more as portions grew (2025 study). More on that in Can ChatGPT count calories from a photo?
- A 2023 systematic review of automated image methods, mostly research systems, found average calorie errors ranging from 0.1% to 38.3%, with simple single foods easiest.
Why photo estimates run low
Three things a camera can’t see well account for most of the gap: fat (oil soaked into food, butter melted into a sauce), portion depth (from above, a deep bowl and a shallow one look alike), and hidden ingredients in mixed dishes. Our guide How accurate are AI calorie apps? explains each one with sources.
Accuracy claims, and what’s behind them
Several apps state an accuracy figure. We report each one as the vendor’s claim, with what we could find about how it was measured.
- Cal AI says on its FAQ that it is “about 80% accurate” (Cal AI FAQ); older posts on its blog say “90% accuracy on visible foods” (Cal AI blog). Neither page says what “accurate” means or how it was tested. In the NIH test, Cal AI’s photo feature had the largest average under-count of the four apps.
- SnapCalorie says it is “on average around 15% mean caloric error” and “2x more accurate than nutritionists” (SnapCalorie FAQ). Its founder co-authored the Nutrition5k paper with Google Research (CVPR 2021), which has a public dataset: a research model missed by 26.1% from a single photo and 16.5% with a depth camera, on dishes from one cafeteria averaging 255 kcal. The paper’s human comparison asked 16 non-nutritionists and 4 nutritionists to estimate portion weights on 10 simple plates (53% and 41% error). That’s a peer-reviewed method, but for a 2021 research model, not a published test of today’s app.
- Fitia calls itself “the fastest & most accurate calorie counter” on the App Store, and Fastic says its scanner “can accurately identify foods”, without a number or method.
- PlateLens, which isn’t in the table above, states “±1.1% kcal error across 180 weighed meals”, citing a study and an open benchmark (PlateLens). We couldn’t check it: the DOI it gives for the study returned “not found” on October 11, 2026, we found no record of the study in Crossref, PubMed or Zenodo, and the benchmark’s repository contained code but no meal photos or reference data.
None of the 12 other apps in our table publishes an accuracy test of its current app with its data, method and date. That’s the standard we’d want before trusting a number, our own included.
What to look for in a photo calorie app
- A list you can check. One total is hard to question; a list of ingredients with portions shows where the estimate went wrong.
- Easy corrections. Changing grams, swapping a food, adding the oil or describing a fix in words should take seconds.
- Barcode and label scanning, so packaged food gets its declared values instead of an estimate. Labels aren’t perfect: in the US a food is only misbranded when it has more than 20% more calories than its label (21 CFR 101.9).
- A way to tell the app what it can’t see, like “cooked in butter” or “half portion”.
- Honesty about accuracy. A claim with a method and data you can check is worth more than a bigger number without one.
How to get better estimates from any app
- Add the fats. Log cooking oil, butter, dressing or mayo as their own item; a tablespoon of olive oil is about 119 kcal.
- Check the biggest items first. Rice, pasta, meat and sauces move the total most.
- Weigh a few staples now and then to calibrate your eye: your usual bowl of cereal, scoop of rice or pour of oil.
- Take a better photo: good light, the whole plate, and a 45-degree angle rather than straight down.
- Use the barcode or label for anything packaged.
Where Calorix stands
Calorix wasn’t part of the NIH test, and we don’t claim it would do better: no camera sees oil soaked into rice. What Calorix does is treat every photo as a first draft. It lists each ingredient with its grams, calories and a confidence badge, so you can see which items deserve a second look, change any of them, add a note like “cooked in butter”, or tell Fix with AI what to correct by text or voice. Barcodes and nutrition labels cover packaged food. See how the Calorix scanner works.
Change log
- October 11, 2026: page published. Features, prices and claims checked on each app’s App Store listing, help center and website; independent studies checked on PubMed, PMC and the publishers’ sites.
Updated October 11, 2026
