White paper · September 2026 · Version 1.0
Meal intelligence.
Measured, not assumed.
How accurately can one photo estimate a meal? We evaluated Fitmo Meal Intelligence on 3,229 weighed Nutrition5k meals and compared it with the FatSecret results reported in an independent publication for the same locked cohort.
Download the white paper PDF ↓01 · The comparison
Less error on the same meal cohort.
Fitmo averaged 84.6 kcal of absolute calorie error, compared with 149.6 kcal for FatSecret. The FatSecret figures come from this independent publication and its public benchmark repository. We compare against those published results, not a new test of FatSecret’s current app.
Calorie error
Mean absolute error (kcal). Lower is better.
Meals within 25% of reference calories
Share of valid results. Higher is better.
| Measure | Fitmo | FatSecret |
|---|---|---|
| Calorie percentage error (MAPE) | 40.2% | 87.9% |
| Portion-mass absolute error | 58.2 g | 92.3 g |
| Portion-mass percentage error | 28.3% | 63.3% |
| Ingredient agreement (Jaccard) | 0.674 | 0.583 |
| Calorie concordance (CCC) | 0.735 | 0.597 |
02 · Beyond calories
Protein, carbohydrates, and fat.
Fitmo also estimated macronutrients directly. The released FatSecret results do not include these outcomes, so this is Fitmo’s accuracy profile, not a head-to-head macro comparison.
Macronutrient absolute error
Mean absolute error in grams. Lower is better.
Macronutrient percentage error
Mean absolute percentage error. Lower is better.
Percentage error can become very large when the true nutrient amount is small. A 4 g error against a 2 g reference is 200%. These are editable estimates, not exact measurements.
03 · Meal complexity
Results vary with what is on the plate.
The study grouped meals by visible ingredient count before scoring. Lower percentage error in more complex meals does not mean every ingredient was easier to identify.
| Visible ingredients | Meals | Calorie MAPE | Mass MAPE | Ingredient agreement |
|---|---|---|---|---|
| 1 | 984 | 51.3% | 34.4% | 0.885 |
| 2–3 | 815 | 38.6% | 26.0% | 0.708 |
| 4–7 | 1066 | 35.8% | 26.7% | 0.546 |
| 8+ | 364 | 26.3% | 21.7% | 0.419 |
04 · Method and context
A locked cohort. A defined comparison.
Each meal supplied one overhead color image and weighed ingredient-level reference data. Fitmo’s output was evaluated without manual correction, database enrichment, or post-hoc calibration. The cohort was locked to the published dish-ID list, with no result-driven sample removal.
Uncertainty was assessed with 10,000 bootstrap resamples. The report includes error, agreement, signed bias, and percentage-within-reference metrics, rather than relying on one accuracy score.
Median inference turnaround was 2.50 seconds; the mean was 2.58 seconds and the 95th percentile was 3.77 seconds. These timings exclude camera interaction, upload overhead outside the timed request, and screen rendering.
05 · Limitations
A useful starting point, not a measurement.
Nutrition5k comes from California cafeterias. It does not represent every cuisine, home meal, beverage, or packaged food. Controlled overhead photos differ from handheld phone images. Hidden oil, sauces, scale, and lighting can change accuracy.
Review and edit meal estimates before saving. These results do not establish clinical benefit and should not guide insulin dosing or replace a dietitian, laboratory measurement, or verified nutrition label.
The next evidence step is a prospectively locked, diverse phone-photo study with weighed ingredients and competing apps tested on identical images.
Read the complete PDF, figures, and references ↗