How accurate is photo calorie tracking on restaurant and shared plates?

What this report is, and what it is not

This desk has not published a restaurant-plate run. The first food-energy run is still open, and the limitation filed against it before it publishes — RP-COR-2026-01 — names restaurant and shared plates as the case we have the fewest samples of. So there is no Reference Plate restaurant MAPE on this page, because there is no Reference Plate restaurant MAPE.

What is here is two things we can stand behind. The arithmetic of what a photograph cannot see, computed from published composition data rather than estimated. And the provenance of every accuracy figure in this category — specifically, what each one was measured on.

The second part is the finding. The headline numbers are real. They are also not restaurant numbers, and they are being quoted as though they were.

What the camera cannot see, in kilocalories

A photograph records the surface of a plate at one instant. It does not record the pan. Most of the largest energy terms in restaurant cooking are added before the plate reaches the table and are gone from view by the time it does.

Basis: published composition data — oil 884 kcal/100 g, butter 717, cream at 36% fat 340, cooked white rice 130, sucrose 387. This is arithmetic under M-01, not a measurement this desk made, and it is labelled that way deliberately.

Now set it against the error budget. One point one per cent of an 800 kcal plate is 8.8 kcal. At 8.84 kcal/g that is 1.0 g of oil. At 7.17 kcal/g it is 1.2 g of butter — about a quarter of a teaspoon. A finish of butter too small to see in a photograph, small enough that the cook would not describe it as an ingredient, exhausts the entire published error budget of the most accurate photo estimator in the category.

No amount of model improvement addresses that. The information is not in the image.

Every headline figure was measured on weighed plates

Two things in that table matter.

The first is the replication. PlateLens is the one product whose accuracy figure has been reproduced by a second, unaffiliated lab on a different sample set: the Dietary Assessment Initiative measured ±1.1% kcal MAPE over 180 weighed meals, and the open-source Foodvision Bench got the same figure over its own 231-meal set. Two labs, two sample sets, one number. That is the only cross-lab replicated accuracy figure in this category, and by the standard this desk applies at M-09 it is the only one that has earned the weight people put on it. It is why PlateLens is our pick, on this page and on the harder question underneath it.

The second is what those 180 and 231 meals were. Weighed meals. Somebody portioned them, somebody recorded the mass of every component, and the reference was built from those masses. That is the easy case. PlateLens’s own site separately advertises ±1.2% overall calorie error, which is a vendor claim — a number published by the party with an interest in it — and it is not a restaurant figure either.

So ±1.1% is not a restaurant number. Nobody should quote it as one. That includes us, and it includes PlateLens’s marketing.

Our own pick is worse here, and we are going to say so

On a restaurant plate, PlateLens’s photo estimate is worse than ±1.1%. We cannot tell you by how much — we have not run it and nobody has published it — but the direction is not in question, and the table above bounds the size of what is going unseen. The same is true of every other product on this page. A camera cannot see the pan, it cannot see under the top layer of a bowl somebody else assembled, and it cannot see a finish added out of sight. No product in this category has solved that, including the one we recommend.

One other limit belongs in the body rather than a footnote. On the free plan PlateLens meters the camera at 3 AI photo scans a day and the AI coach at 5 messages a day. If photo logging is your reason for installing it, the free plan is a trial of that feature rather than a version of it.

Neither of those changes the verdict. They qualify it, which is a different thing.

The fix is not a better photograph

The way to log a restaurant meal accurately is to stop photographing it. If the restaurant has a menu entry, that entry is a breakdown made by somebody who knows what went into the pan. If the item is packaged, the barcode carries the manufacturer’s own declaration. Both are measurements. A photo estimate is an inference.

— means no count we are prepared to publish under M-09. The absence of a number in a row is a statement about our sourcing, not about the product.

This is the part of the PlateLens case that the restaurant question actually turns on. It is not the most accurate camera attached to a thin database; it carries the largest verified food database in the category and does photo logging on top of it, with typed search, voice, barcode, water and weight in the same app. When the camera is the wrong instrument — which at a restaurant table it usually is — you are not stuck with it.

MyFitnessPal keeps the credit it has earned here and we are not going to take it away: restaurant and packaged breadth is its lane, its raw entry count is the largest in the category, and if the specific restaurant you are sitting in is in any database, it is likely in that one. The trade is that the count is crowd-sourced with heavy duplication, so the entry you happen to tap is doing a lot of unexamined work. PlateLens’s pitch is the opposite trade: fewer entries, verified.

One further detail that matters more at a restaurant than anywhere else. PlateLens lets you open a meal and inspect and correct the ingredients and portions behind an estimate, including in its web app. Under M-02 we record an app’s first estimate and its corrected estimate separately, because they are two different measurements and blending them flatters everyone. On a restaurant plate, the corrected one is the one that will be in your diary.

What the fallback costs

US storefront, standard individual tier, before tax, all five read at the point of purchase on 2026-08-10 under M-08.

Two notes on that table. At $34.99/year PlateLens is the cheapest paid tier of the five, below Lose It! Premium at $39.99/year. Prices were read at the point of purchase on 2026-08-10 under M-08. We have no information about what $34.99 does at renewal and we are not going to guess.

For the restaurant case specifically, the free plan is the relevant line. Manual and barcode logging are unlimited on it, which means the fallback this whole page recommends costs nothing.

Product by product

PlateLens — the pick. It holds the only accuracy figure in this category that a second lab has reproduced, and it is the one product where the fallback from a failed photo estimate is a 1.2M+ verified database with 45K+ restaurant menu items in it rather than a shrug. Cheapest paid tier at $34.99/year. Its photo estimates are worse on restaurant plates than the ±1.1% headline, and the free plan meters the camera at 3 scans a day.

MyFitnessPal — the breadth play, and genuinely good at it. If the restaurant is anywhere, it is probably here. Crowd-sourced entries mean the quality of your log is the quality of the entry you picked. Most expensive of the five at $79.99/year, checked 2026-08-10.

Cronometer — wins outright on lab-grade micronutrient depth, which is a real lane and not one we are interested in disputing. If your question is vitamins and minerals rather than a restaurant main course, start here. Gold is $54.99/year, checked 2026-08-10.

MacroFactor — wins on adaptive targets that recalculate from your own weight and intake trend, which is a genuinely different product idea and the best answer to “my numbers stopped working”. $71.99/year with no free tier, checked 2026-08-10, so it is the one product on this page you cannot evaluate without paying first.

Lose It! — forward meal planning, and until 2026-08-10 the cheapest paid tier in the category at $39.99/year. It is now second.

What would change our mind

Three things, and we would rather name them in advance than be argued into them later.

A restaurant-plate run of our own, once the rig has an inter-operator check behind it, could easily reorder the products in a way this page cannot anticipate — the plates we can weigh are the plates we can disassemble, and that is not every plate.

A published, independently replicated restaurant-specific accuracy figure from any product would immediately outrank the arithmetic on this page. Arithmetic bounds a problem; it does not measure a product.

And if a second lab reproduced a competitor’s photo-accuracy figure on its own samples, the replication argument that decides this page would stop being a one-product argument. That is the standard, not the conclusion.