Benchmarks
How Truelens compares with calling the model yourself
Every API was run on held-out photos whose answers were measured in a lab or graded by experts, next to one plain call to the same model given the same photo and the same task. This page explains how each test works and lists every number we recorded.
2x
more meals within 20% of the true calories
60 real meals, each weighed and measured in a lab
Calories within 20%
2xCalories within 30%
+18 ptsPortion weight within 20%
2.6x| Measure | Truelens | Plain model call |
|---|---|---|
| Calories within 20% | 40% | 20% |
| Calories within 30% | 53% | 35% |
| Portion weight within 20% | 46% | 18% |
99%
of acne grades within one step of a dermatologist
150 faces graded by dermatologists
Grade exactly right
+17 ptsGrade within one step
+11 pts| Measure | Truelens | Plain model call |
|---|---|---|
| Grade exactly right | 55% | 38% |
| Grade within one step | 99% | 89% |
93%
of Nutri-Score grades exactly right
59 real supermarket label photos with published grades
Nutri-Score exactly right
+29 pts| Measure | Truelens | Plain model call |
|---|---|---|
| Nutri-Score exactly right | 93% | 64% |
88%
of species found in the top 3 answers
91 phone photos of 25 species, identified by expert reviewers
Species in top 3
+3 ptsSpecies as first answer
+3 pts| Measure | Truelens | Plain model call |
|---|---|---|
| Species in top 3 | 88% | 85% |
| Species as first answer | 77% | 74% |
How the tests work
Answers we did not make up
Each test set comes with answers from outside: meals weighed and measured in a lab, faces graded by dermatologists, Nutri-Score grades published for real products, and plants identified by expert reviewers.
Photos kept out of tuning
For food, the 60 test meals were set aside before tuning, which used 40 other meals, and were scored once at the end. For skin and plant, two choices were made after seeing results on the test photos (grading acne from lesion counts, and the model used for plants), so treat those two as slightly optimistic until the next fresh test set.
A fair plain call
The comparison is what a developer gets from one well-written call to the same model: the same photo, the same scale or question, and a structured answer. It is not a weak prompt chosen to lose.
Same model on both sides
Food, skin and label use Claude Haiku 5.5 for both methods; plant uses Claude Sonnet 5.5 for both. Any gap comes from how Truelens works, not from a bigger model. Caching was off, so every answer was a fresh call.
Food & calories
What it was tested on
60 real cafeteria meals, photographed from above, each weighed and measured in a lab so the true calories, weight and macros are known. Every meal was scanned twice by each method, so each figure covers 120 answers.
How it was scored
A meal counts as right when the estimate is within 20% (or 30%) of the measured calories. We also report the average error against the measured values, the measure nutrition studies usually use.
| Measure | Truelens | Plain model call |
|---|---|---|
| Calories within 10% | 18% | 12% |
| Calories within 20% | 40% | 20% |
| Calories within 30% | 53% | 35% |
| Portion weight within 20% | 46% | 18% |
| Average calorie error | 40% | 96% |
| Average portion weight error | 34% | 79% |
| Average protein error | 70% | 102% |
| Average fat error | 79% | 141% |
| Average carbohydrate error | 196% | 236% |
- Carbohydrate error is high for both methods because many of these meals have almost no carbs, so a small miss becomes a large percentage.
- The photos are overhead cafeteria shots. Phone photos taken at an angle may score differently.
Skin analysis
What it was tested on
150 face photos with acne of every severity, each graded by dermatologists on a four-level severity scale based on lesion counts.
How it was scored
We compare the acne grade with the dermatologists’ grade. The plain call was given the same four-level scale and the same lesion-count bands, so both sides answer the same question.
| Measure | Truelens | Plain model call |
|---|---|---|
| Grade exactly right | 55% | 38% |
| Grade within one step | 99% | 89% |
| Agreement with dermatologists (weighted kappa, 0 to 1) | 0.67 | 0.57 |
| Average lesion count error | 8.0 lesions | 18.9 lesions |
- The grade comes from the lesion counts with a fixed rule, which matched dermatologists better than asking the model for a grade directly.
- Other skin concerns (wrinkles, redness, pores and so on) have no public labelled test set yet and are not part of this benchmark.
Product label
What it was tested on
59 real supermarket product photos of nutrition panels, each with the Nutri-Score grade published for that product.
How it was scored
A product counts as right when the grade matches the published one exactly. The plain call read the same photo and was asked for the grade under the current (2023) algorithm.
| Measure | Truelens | Plain model call |
|---|---|---|
| Nutri-Score grade exactly right | 93% (55 of 59) | 64% (38 of 59) |
| Grade within one step | 95% | 95% |
| Nutrient values read from the photo, average error | 3% | 3% |
- Both methods read the numbers on the panel equally well. The difference is the grade: Truelens computes it with the official formula instead of letting the model estimate it.
- Two of the four misses were rotated or blurred photos. Truelens declined to grade them and said why; the plain call guessed.
- We separately checked the grading rules against 106 published products: the computed Nutri-Score matched in every case, the processing level (NOVA) in 95%, and allergens were found with 95% precision and 98% recall.
Plant ID & health
What it was tested on
91 phone photos of 25 common houseplants and garden plants, each identified by expert reviewers. Photos were sent at about 1000 pixels, close to what an app uploads.
How it was scored
We check whether the true species is the first answer, or among the first three. Names are resolved to their accepted scientific name first, so synonyms count as correct for both sides.
| Measure | Truelens | Plain model call |
|---|---|---|
| Species as the first answer | 77% | 74% |
| Species in the top 3 answers | 88% | 85% |
| Genus as the first answer | 91% | 91% |
- Plant identification is where a strong model already does well, so the gain here is smaller. What Truelens adds on top is the checked name, pet toxicity, care and treatment steps for each health problem.
- Most misses are between close relatives, such as two oak species or two Philodendrons, where even the leaves of the true species vary a lot.
Not benchmarked yet
The body and color APIs have no public test set with trustworthy labels that may be used to evaluate a commercial product. Body fat needs photos paired with a lab measurement such as a DEXA scan, and color seasons need portraits graded by trained color analysts. We will publish both once we have such a set. Until then, both were checked for consistent answers on repeated photos, and in production an identical photo always returns the same result.
Versions
Results as of October 2026. This page is updated whenever the benchmarks are rerun.