Accuracy, published

These are the real numbers from Renzy’s scanner against a set of reference dishes, no marketing: the average error in calories and in grams, how it is spread and the worst cases, recalculated every Monday.

MAPE on kcal

34%

On average, how far the estimated kcal are from each dish’s reference value.

MAE on kcal

On average, how many kcal off the reference value. Lower is better.

MAPE on grams

46.8%

On average, how far the estimated amount, in grams, is from the reference amount.

In this batch the estimate was too high on 6 dishes and too low on 2.

How it's computed

Each dish in the set is a photo with a reference value. There are only a few (the sample size is shown below), and almost all of them are stock photos (Unsplash): branded products with the grams and calories on their label, and simple foods with a standard USDA or BEDCA portion. They are not dishes weighed on a scale: nobody weighed what is in each photo, and if a photo shows more or less food than the reference, that gap also counts as scanner error. So the figure shows whether the scanner gets better or worse from one week to the next on the same photos, not the error you’ll get on your own meals.

Each photo carries its reference value and where it comes from (manufacturer label, USDA or BEDCA). The worst cases are shown below, with their photo.
What is measured is the app’s first analysis: the same AI model and the same automatic corrections (checks against the food database, sauces, brands, garnishes), with no extra context and no cache. It leaves out the extra passes the app runs afterwards on some dishes (doubtful portions, drinks, a second model) and what it learns from your corrections.
The average error is the mean of each dish’s error, with none left out. With so few dishes, one big miss moves the average a lot: that is why the spread and the worst cases are shown below.
It reruns automatically every Monday, on the photos in the set. Whatever comes out is published as is, including when it gets worse.

Error distribution

From the dataset's plates, this is the distribution of percent error on kcal. The tighter the bars cluster on the left, the better.

0–5%

5–10%

10–20%

20–30%

30–50%

50%+

The 5 worst cases

What came out worst in this batch, with its photo and its reference value.

Yogur griego natural Hacendado

Reference: 145 kcal · Estimated: 437 kcal

Error: 201.4%

Tostada con aguacate y huevo poché

Reference: 380 kcal · Estimated: 277 kcal

Error: 27.1%

Caña de cerveza 200ml

Reference: 86 kcal · Estimated: 104 kcal

Error: 20.9%

Pasta carbonara 320g

Reference: 560 kcal · Estimated: 604 kcal

Error: 7.9%

Ensalada César con pollo 280g

Reference: 470 kcal · Estimated: 504 kcal

Error: 7.2%

Last run: November 2, 2026

Sample size: 8 platos

Model: anthropic/claude-opus-5.5

Batch: cron-2026-11-02-mvyvqfx0

Want to replicate the benchmark with your own dataset? Email us at hola@renzy.app.