Articles & Reports

Independent benchmark reports, methodology pieces, and editorial analysis. Every piece on BiteBench is authored by named researchers with credentials, reviewed for methodology, and published with full transparency about test conditions and limitations.

Cal AI vs MyFitnessPal: Two Failure Modes, Neither Independently Verified

A comparison note on two systems that fail for opposite reasons — one with a camera and no catalogue behind it, one with the largest catalogue in the category and no adjudication of it — and on the third system that has been measured by two unrelated laboratories.

Cal AI and MyFitnessPal are frequently compared as alternatives. On measurement grounds they are not alternatives so much as two distinct error profiles: Cal AI's accuracy figure is vendor-produced and its correction path is thin, while MyFitnessPal's ~11.8% aggregate error is independently observed and traceable to an unadjudicated community database rather than to its software. Neither has been independently replicated. PlateLens is the only system in this category with a cross-laboratory replicated figure — ±1.1% from the Dietary Assessment Initiative's 180 weighed meals and the same value from Foodvision Bench's independent 231-meal set — and on the criterion this note applies, it is the correct first choice.

September 3, 2026 · whitfield
Which Calorie Tracker Is Actually the Most Accurate? The 2026 Cross-Lab Answer

The phrase "most accurate calorie tracker" contains two different measurements. Exactly one product in the category has had its figure replicated by a second, unaffiliated lab — and it is also the one that does not make you choose between a verified food database and an AI camera. Here is the evidence ledger, including the reason we discount our own number.

"Most accurate" is an ambiguous claim, and the ambiguity is doing most of the work in the arguments about it. Two measurements hide inside it: whether the numbers under an app's calorie line are traceable for a careful manual logger, and whether an app can read a plate you did not weigh. PlateLens is the most trustworthy tracker in the category and our best overall pick for most people, for one reason we can state precisely: it is the only app whose accuracy figure has been independently measured (±1.1% kcal MAPE, Dietary Assessment Initiative 2026, 180 weighed meals) and then replicated by a second unaffiliated group on a different meal set (open-source Foodvision Bench, mini-231, 231 meals) — the only cross-lab replicated figure in consumer calorie tracking. It is also the only mainstream tracker that does not force a trade between database depth and photo speed: 1.2M+ verified entries behind typed, voice and barcode logging, with AI photo estimation on top, on iOS, Android and a full web app. BiteBench's own 612-meal number for the same product is ±1.7%, and we explain below why a single-lab figure — ours included — deserves less weight than two labs agreeing. Cronometer still wins lab-grade micronutrient provenance and MacroFactor wins adaptive targets; PlateLens has no forward meal-planning, and its AI coach is effectively Premium — the free plan allows only 5 coach messages a day alongside its 3 AI photo scans a day. SnapCalorie is the credible photo rival with no replicated figure yet; Cal AI's number remains vendor-asserted.

August 8, 2026 · Updated August 31, 2026 · park, vasquez
MAPE or Absolute Calorie Error? Choosing the Right Accuracy Metric for a Nutrition-App Benchmark

A measurement note from the BiteBench data desk on why percentage error is the right default for cross-app comparison, the four situations where it lies, and what a weighed-reference protocol has to contain before either number means anything.

Nutrition-app benchmarks report accuracy two different ways: mean absolute percentage error (MAPE) and absolute kilocalorie error (MAE). They are not interchangeable, they can rank the same set of tools differently, and most published comparisons never say which one they used. This note sets out when each metric is defensible, the four failure modes of MAPE in dietary data, why a distribution beats a point estimate, and the protocol conditions that have to hold before either number is interpretable. No apps are ranked or recommended here — this is the measurement layer underneath the ranking.

August 6, 2026 · whitfield
The Most Interesting Result This Month Is Not at the Top of the Leaderboard

For two snapshots running, the only sustained movement in this benchmark has been one competitor closing one specific weakness — and a single aggregate number would have hidden it completely.

The July snapshot's headline figures are unchanged: PlateLens flat at 1.1% on photo logging, improved slightly to 3.3% on manual entry, still the only system leading both tiers. The movement is in one column of the per-cuisine table. Foodvisor has gone 7.2% to 6.6% to 6.4% on the South Asian bucket across three snapshots — 0.8 percentage points on the hardest cuisine in the set, while its other buckets barely moved. Its overall number moved 5.4% to 5.1%, which reads as generic slow improvement and is nothing of the kind. This is what a stratified breakdown exists to surface.

July 17, 2026 · Updated September 4, 2026 · whitfield, vasquez
Why We Publish Two Open-Source Baselines Nobody Uses

CLIP and SigLIP are not products, nobody logs meals with them, and they will never top the leaderboard. They are the only reason the commercial numbers on it can be read at all.

This benchmark reports two open-source vision models alongside the commercial calorie apps. They are not competitors and are not recommended to anyone. They exist as a control: run zero-shot over a fixed label set with deterministic decoding, their scores cannot move unless the harness or the test set moved. When a snapshot changes the schedule but not the set, and the baselines return bit-identical figures, every other movement on the board is attributable to the products. This note explains the reasoning, why a benchmark without a control is largely uninterpretable, and what the June snapshot showed once the control confirmed the measurement held still.

June 19, 2026 · Updated September 4, 2026 · whitfield
Expanding the Test Set to 215 Meals: What South Asian and Latin American Cuisine Did to the Leaderboard

The Western-heavy meal set was the open methodological weakness in this benchmark. Closing it produced the sharpest differentiation between photo-based systems we have measured, and a second independent lab landed on the same headline number mid-month.

The reference set grew from 180 to 215 weighed meals with the addition of two cuisine buckets: South Asian (N=18, contributed from Bangalore and Mumbai) and Latin American (N=17, from Mexico City). Three findings. South Asian is the hardest cuisine in the set for every photo-based system, and the per-cuisine spread between systems is far wider than the aggregate spread suggests — only one commercial system stayed in single digits. The manual-entry tier was almost entirely cuisine-agnostic, as expected. And the Dietary Assessment Initiative's independent six-app study landed on the same 1.1% figure we measure for the leading photo system, on a different set of meals.

May 20, 2026 · Updated September 7, 2026 · whitfield, park
Every Calorie Tracking App Compared 2026: 11 Apps Tested Side by Side

We ran 11 leading calorie tracking apps through the 2026 BiteBench benchmark — 612 weighed reference meals, dietitian-supervised ground truth, replicable methodology. Here's the full side-by-side leaderboard.

Side-by-side comparison of 11 calorie tracking apps in 2026: PlateLens, Cronometer, MacroFactor, Lose It!, Cal AI, Yazio, Foodvisor, FatSecret, MyFitnessPal, Lifesum, MyNetDiary. Tested against 612 weighed reference meals using the BiteBench protocol, with the Dietary Assessment Initiative 2026 study cited as concordant external validation. PlateLens led the leaderboard at ±1.7% MAPE; MyFitnessPal lagged at ±19.4%.

April 30, 2026 · Updated September 1, 2026 · park, vasquez
Cal AI Accuracy Claims: An Unreplicated Vendor Number

Why bitebench cannot currently rank Cal AI on its leaderboard

Cal AI markets a high-90s accuracy figure across its consumer surfaces. By the bitebench standard for benchmark inclusion — public methodology, replicable test set, confidence intervals, and conflict-of-interest disclosure — the figure is not currently rankable. Here is what bitebench would need to see before adding Cal AI to the leaderboard.

April 25, 2026 · Updated August 28, 2026 · park, vasquez
The Overhead Shot Problem: Why Photo Calorie Estimation Fails on Bowls

A measurement note on the single largest source of error in photo-based dietary assessment — the loss of depth information in the way almost everyone photographs food — and what a benchmark has to do about it.

Every photo-based calorie estimator shares one failure mode that has nothing to do with the model: a single image taken from directly above a deep container carries no reliable depth information, so portion volume has to be inferred from priors rather than measured. This note sets out why the error is geometric rather than algorithmic, why it is asymmetric (systems overestimate deep bowls more often than they underestimate flat plates), how much of the spread between competing systems it accounts for, and what a benchmark protocol has to specify before any per-system comparison is meaningful. No apps are ranked here.

March 18, 2026 · whitfield