Cal AI、MyFitnessPal、Lose It! 等四款照片卡路里應用程式測試102頓餐點，估算值全數偏低

If you are building anything that turns a photo of food into a number, this one is worth a minute.

Researchers in the intramural program of the US National Institute of Diabetes and Digestive and Kidney Diseases (NIH) cooked 102 meals in a metabolic kitchen, where the content of every plate is known. They analysed photos of each meal with the photo features of four apps and compared every estimate with the known value.

Every app came in low. Mean energy underestimate per meal:

That is roughly a third of the meal missing. All four also underestimated fat, by about 30 g; Lose It! and Cal AI underestimated carbohydrate too, by about 14 g.

The fat line is the interesting one

Thirty grams of fat is about 270 kcal on its own (9 kcal per gram, the standard Atwater factor). That is close to the whole gap. Our reading, not the study's: fat is the part of a plate a camera sees worst — oil in the pan, butter melted into the rice, the dressing under the leaves. A vision model can name the salad and still miss the tablespoon of oil.

You can see how much of a food's energy sits in fat with any open nutrient table. Here is one, the FatTonic Food Index, where every row points at the USDA record it was read from:

import csv, io, urllib.request

url = "https://fattonic.com/data/export/foods.csv"
foods = list(csv.DictReader(io.TextIOWrapper(urllib.request.urlopen(url), "utf-8")))

def kcal(f):  # Atwater factors: 4 kcal/g carbohydrate and protein, 9 kcal/g fat
    return 4 * float(f["carbs_g"]) + 4 * float(f["protein_g"]) + 9 * float(f["fat_g"])

for f in sorted(foods, key=lambda f: 9 * float(f["fat_g"]) / max(kcal(f), 1), reverse=True)[:5]:
    share = 9 * float(f["fat_g"]) / kcal(f)
    print(f'{f["name"]:<22} {f["serving"]:<18} {share:4.0%} of energy from fat  {f["source_url"]}')

Run today it prints pecans (87 %), walnuts (84 %), hazelnuts (81 %), avocado (75 %) and cheddar (73 %), each with its USDA FoodData Central link. A handful of nuts on a photo is a small object carrying most of its energy as fat, which is where a size guess would hurt most.

What the study does not show yet

It was presented at the American Society for Nutrition's NUTRITION meeting as a conference abstract, not a peer-reviewed paper, and the report calls the results preliminary until the full study is reviewed. It measured accuracy, not whether using an app changes anyone's health. The lead researcher's advice for a photo logged without checking portions: take the result "with a grain of salt".

The design choice it points to

FatTonic's camera does not estimate a mixed plate. It names one fresh food at a time, you tap the match, and the number comes from that food's sourced row. It is a narrower feature. This study is a good argument for keeping it narrow.

• The study, the table and the caveats: How accurate are AI photo calorie apps?
• The Food Index data (CSV, JSON): fattonic.com/data
• Where your own diet stands: the free Metabolic Score

Source: Healio, Emma Bascom, "AI photo-based calorie-tracking tools underestimate them by 33%", 4 August 2026. Information, not medical advice.