Our Benchmark Methodology

Scoring Framework

Each app is scored across 5 AI-specific categories with the following weights:

Category	Weight	Measured by
Recognition Accuracy	30%	Food identification rate (% correct)
Portion Estimation	25%	Portion MAPE vs dietitian-weighed values
Speed	20%	Median time from shutter to diary entry
Food Category Coverage	15%	Recognizable food categories / cuisine breadth
Learning & Adaptation	10%	ID rate improvement after correction feedback

Each category is scored 0–10. The weighted sum produces the overall score out of 10.

Image Acquisition Protocol

All test images were captured on an iPhone 15 Pro rear camera (48MP main sensor, 1x zoom) with the following standardized conditions:

• Lighting: Two 5500K LED panels providing 120 lux incident illumination at food surface level, positioned at 45-degree angles to minimize harsh shadows
• Distance: 30cm from lens to food surface (measured with positioning jig)
• Angle: Approximately 60 degrees from vertical (typical overhead phone hold)
• Plate: White circular ceramic plate, 27cm diameter, standard across all American/European food presentations
• Background: Neutral gray matte surface
• Image resolution: Native camera output, not resized before submission to apps

Ground Truth Preparation

Each meal composition in the test set was:

1. Prepared with precise portion weights measured on a calibrated digital scale (resolution: 0.1g)
2. Photographed before any disturbance
3. Cross-referenced against USDA FoodData Central for calorie and macronutrient values
4. Verified by our registered dietitian consultant for nutritional accuracy
5. Logged with a ground truth calorie value and food category label per item

12 meal compositions were excluded from the portion MAPE analysis because ground truth calorie uncertainty exceeded ±3% (due to highly variable-density foods like handmade doughs). These meals were retained in the identification rate analysis.

Identification Rate Measurement

A food item is counted as "correctly identified" if the app's top-1 prediction matches the ground truth food category at the dish level. Example: "grilled salmon" is correct for a grilled salmon fillet, but "fish" alone is not. "Spaghetti bolognese" is correct; "pasta with sauce" is not.

For multi-item plates (e.g., a plate of chicken, rice, and broccoli), each food item is scored separately. The identification rate is the proportion of all individual food items across all 500 images that were correctly identified.

Apps that return a "cannot identify" or "low confidence" result rather than a wrong guess are not counted as incorrect — these are excluded from the denominator. This rewards apps that use confidence thresholds rather than forcing a guess.

Portion MAPE Measurement

Portion MAPE (Mean Absolute Percentage Error) is calculated as:

MAPE = (1/n) × Σ(|app_kcal − true_kcal| / true_kcal) × 100

Where n is the number of images in the portion accuracy test set (n=488 after exclusions), app_kcal is the calorie value reported by the app for the complete meal, and true_kcal is the ground truth calorie value derived from the weighed portions and USDA reference.

Processing Speed Measurement

Processing speed is measured as the median of 50 repeated trials per app:

• Timer starts at camera shutter release
• Timer stops when the food diary shows a completed entry (food identified and logged, user able to interact with the result)
• Measured on iPhone 15 Pro
• Wi-Fi connection: 100 Mbps down / 50 Mbps up, measured before each session
• App freshly opened for each trial (cold start) to control for caching
• Median used rather than mean to reduce influence of outlier delays

Food Category Coverage Assessment

We attempted to log 100 challenging food items representing rare or regional foods not commonly found in US grocery stores. The category coverage score reflects both the count of supported food categories and the breadth of cuisine types represented. Apps were also penalized if they repeatedly returned "cannot identify" on common foods from non-Western cuisines.

Learning & Adaptation Assessment

We created fresh accounts in each app, logged 200 corrections over a 30-day period for foods the app initially misidentified, and retested identification rate on the same 100 misidentified images at day 30. The improvement in identification rate (and the quality of the correction interface) determines the learning & adaptation score.

Independence and Conflicts of Interest

We have no commercial relationship with any app reviewed on this site. We did not receive press access, free subscriptions, or payment from any app developer. All apps were tested using standard consumer accounts. Some outbound links on this site are affiliate links; they do not influence our scores or rankings in any way.

Alex Park (author) has no financial interest in any reviewed company. Dr. Kenji Yamamoto, PhD (technical reviewer) has no financial interest in any reviewed company. Dr. Yamamoto reviewed the benchmark methodology and statistical analysis; he did not conduct the individual app tests.

Data Sources

• USDA FoodData Central: fdc.nal.usda.gov — primary nutritional reference for ground truth calorie values
• NCCDB: Nutrient Coordinating Center Database — secondary reference for restaurant items and prepared foods
• App-specific databases: Compared against USDA values to assess nutritional data quality

Retest Cadence

We retest all apps on the same benchmark dataset every 6 months, or immediately following a major model update announcement from any tested app. Scores shown reflect the most recent benchmark run (March 2026).