The honest answer, before anything else: AI calorie counters are good at recognizing food and mediocre at guessing how much of it is on the plate.Anyone telling you a single photo gives laboratory-grade numbers is selling something. Here's what the research actually shows — and how to get useful numbers anyway.
What the research says
Published evaluations of photo-based food logging cluster around a few consistent findings:
- Food identification works well.Modern vision models identify what's on the plate correctly most of the time — benchmark reviews put identification accuracy roughly in the 70–90% range for everyday meals, higher for distinct single foods.
- Portion size is the weak spot.A single 2D photo carries roughly 15–25% portion error, and dense foods like rice or pasta can be off by more — the camera can't see how deep the pile goes.
- Photos still beat memory.A 2024 study in JMIR found image-based dietary methods underreport energy intake by about 20% against the clinical gold standard — while traditional written food records underreport by anywhere from 11% to 41%. The photo doesn't forget the handful of fries you did.
Why speed beats precision for real results
The dirty secret of calorie tracking is that the biggest source of error is not the estimate — it's the meals that never get logged at all. Retention research points the same way: people who can log a meal in under 30 seconds stay consistent for months, while people grinding through database searches and portion menus mostly quit within weeks. An 85%-accurate log you keep every day tells you far more than a 95%-accurate log you abandon by February.
This is the design bet behind photo-first tracking: accept a portion estimate you can correct in two taps, in exchange for logging that actually happens.
How LensPlate handles the accuracy problem
Three things, each aimed at a specific failure mode:
- Per-ingredient breakdowns, editable before you log. The AI itemizes the plate — salmon, rice, avocado — with grams for each. If it read 150g of rice and you know it's 250g, you adjust that one slider and the macros rescale. You correct the part the camera can't know, instead of accepting or rejecting a black-box total.
- Barcodes and labels are exact, so use them when you can. For packaged food, LensPlate scans the barcode or reads the printed nutrition label directly — manufacturer data, zero estimation. The camera is for the plates that don't have labels.
- Trends over single meals. Your daily plan and progress ring work on totals and trends, where per-meal noise averages out. Portion errors cut both ways; the weekly picture is much more stable than any one lunch.
Five habits that make photo estimates better
- Shoot from a slight angle (30–45°), not directly overhead — depth cues help portion estimation.
- Keep a fork or hand at the edge of the frame as a scale reference.
- Photograph before you start eating, while everything is visible.
- Split very mixed dishes: sauce-heavy bowls hide ingredients; log the obvious extras (oil, dressing) separately.
- Correct portions when you know better — home-cooked meals you make weekly deserve one accurate edit that you reuse.
The verdict
AI calorie counting is accurate enough for the thing calorie counting is actually for: knowing whether you're roughly on plan, every day, with almost zero effort. It is not accurate enough to treat any single meal's number as truth — which is exactly why the estimate should be editable, and why packaged food should come from barcodes and labels rather than the camera. Tools that hide the uncertainty are doing you a disservice; tools that let you correct it are doing their job.