# How to Photo-Log Calories Accurately (and What the Accuracy Studies Say)
Source: https://www.trackpep.com/blog/how-to-photo-log-calories-accurately
> Educational information only. Not medical advice, and not a dosing recommendation. TrackPep records what you enter; it never calculates, converts or decides a dose for you. Talk to a licensed clinician.
Published error ranges for AI food-photo estimation, why portion size — not food recognition — is the weak link, and the shooting habits that measurably narrow the gap.
## Short answer
A 2023 systematic review found AI calorie estimates from photos averaged errors of 0.10 to 38.3 percent, with individual errors up to 79.6 percent. Single-food images beat multi-food plates. Shoot at an angle, include a size reference, separate mixed foods, and correct the portion the app guesses.
## What "accurate" means here
**Literature checked 3 September 2026.**
The honest answer to "how accurate is a food photo?" is a range, and anyone quoting a single percentage is selling something.
Shonkoff and colleagues' systematic review in *Annals of Medicine* (2023) compared AI image-based dietary assessment against ground truth across the published studies:
| Measure | Average relative error | Individual relative error |
| --- | --- | --- |
| Calories | 0.10% – 38.3% | 0.00% – 79.6% |
| Volume | 0.09% – 33% | 0.00% – 34.2% |
Their stated conclusion is worth quoting rather than paraphrasing: the relative errors "suggest that AI methods align with – and have the potential to exceed – accuracy of human estimations." That is a real result, and also a low bar. People are famously bad at estimating what they ate.
The finding that changes behaviour, though, is this one: **average errors and ranges were smaller for single-food images than for multi-food images.** A photo of one grilled chicken breast is a tractable problem. A photo of a curry with rice, a naan, and half a salad is four estimation problems stacked on top of each other, and the errors do not cancel.
**TrackPep does not give nutritional advice.** Meal Scan estimates what is on a plate so you have something to log and correct. It is not a diagnostic tool, not a substitute for a dietitian, and the estimate is an estimate.
## Why portions, not recognition, are the weak link
Rouhafzay and colleagues' 2025 scoping review in *Frontiers in Nutrition* looked at 14 calorie-counting applications and separated the two jobs an app is doing.
Food **identification** has largely been solved for common foods — several systems reported top-5 classification accuracy above 90%. Portion and volume **estimation** has not. The review's list of what the most accurate volume methods require is the tell: **specialised hardware, a reference object in frame, or multiple images**. Those requirements are exactly what a single casual snapshot lacks.
The review also flags dataset bias — training sets skew toward particular cuisines and regions, so a model's accuracy on your food may not resemble its accuracy on the benchmark. If you eat food that is under-represented in Western food-image datasets, expect the estimate to be worse and correct it more aggressively.
## Six habits that narrow the gap
None of these are exotic. They exist because they attack the two documented failure modes — hidden volume and stacked foods.
1. **Shoot from about 45 degrees, not straight down.** A top-down photo shows area and hides depth. Depth is the whole of the volume problem.
2. **Put a known object in frame.** A fork, a standard mug, your thumb. A reference object is one of the three things the literature says accurate volume estimation needs, and it is the only one you can supply for free.
3. **Separate mixed foods before shooting** when it is not absurd to do so. Single-food images have measurably smaller errors. Pushing the rice away from the curry for two seconds is a real intervention.
4. **Photograph the leftovers.** What you served is not what you ate. A second photo of the plate at the end converts a guess about the portion into a subtraction.
5. **Shoot in even light.** Shadow reads as edge, and edges are what the volume estimate is built from.
6. **Correct the number.** Every one of these studies measures the raw model output. An app that lets you adjust the portion afterwards is being evaluated with the correction turned off. Turning it on is free accuracy.
## Consistency beats precision
Most people photo-logging calories are not auditing a metabolic ward. They are trying to answer "am I eating more or less than last month?" For that question, a **stable bias is nearly harmless and a random error is corrosive**. An app that reads your usual breakfast 15% high every single day still shows you the trend correctly. An app that swings between 20% low and 30% high depending on the lighting does not.
That is why the habits above are worth more than app choice. A repeatable method has a repeatable error.
## Where this bites hardest on a GLP-1
If you are tracking protein — and the 2025 survey data suggest most people on GLP-1 receptor agonists fall short of the published range, with only 43% reaching 1.2 g/kg/day — then portion error lands directly on the number you care about. Protein is concentrated in the dense, hard-to-estimate items on the plate. The meat is where the error lives.
For meals where the protein number actually matters, weigh it. Photo-log the rest.
## What TrackPep does
Meal Scan sends the photo you choose to take for estimation, returns editable items with portions, and logs the totals. **You can correct every field**, and the corrected number is what gets stored. We publish this accuracy discussion rather than a marketing claim because the literature does not support a marketing claim.
## Frequently asked questions
### How accurate is AI calorie estimation from a photo?
Shonkoff and colleagues' 2023 systematic review in Annals of Medicine reported average relative errors for calorie estimation of 0.10 to 38.3 percent, and individual relative errors from 0.00 to 79.6 percent. Volume estimation averaged 0.09 to 33 percent error. The range is the finding — accuracy is not one number.
### What makes photo estimates go wrong most often?
Portion size, not food identification. The 2023 review found errors were smaller for single-food images than for multi-food images, and a 2025 Frontiers in Nutrition scoping review concluded the most accurate volume methods need reference objects, multiple images, or special hardware.
### Is photo logging good enough to be worth doing?
The 2023 review concluded AI estimates align with, and can potentially exceed, the accuracy of human estimation. Humans are not a high bar — people routinely misjudge their own intake — so a consistent method with a known error beats an inconsistent one with an unknown error.
### Does it matter if the estimate is off by the same amount every day?
A consistent bias still shows the direction of a trend, which is what most people are actually using the number for. Errors that swing at random between meals are the bigger problem, and that is what a repeatable shooting habit reduces.
### Does TrackPep photo logging replace weighing food?
No. A kitchen scale is more accurate than any photo method, and TrackPep lets you correct a scanned portion by hand. Photo logging is a convenience for meals you would otherwise not log at all.
## Sources
- [Shonkoff E, Cara KC, Pei X, et al. AI-based digital image dietary assessment methods compared to humans and ground truth: a systematic review. Ann Med. 2023;55(2):2273497.](https://pmc.ncbi.nlm.nih.gov/articles/PMC10836267/)
- [Rouhafzay A, Rouhafzay G, Jbilou J. Image-based food monitoring and dietary management for patients living with diabetes: a scoping review of calorie counting applications. Front Nutr. 2025;12:1501946.](https://www.frontiersin.org/journals/nutrition/articles/10.3389/fnut.2025.1501946/full)
- [Johnson B, McGlasson T, Thomas O, Kreider R, Jones R. Suboptimal protein intake for hypocaloric diet needs while using glucagon-like peptide-1 receptor agonists. J Int Soc Sports Nutr. 2025;22(Suppl 2):2550139.](https://pmc.ncbi.nlm.nih.gov/articles/PMC12419545/)
By TrackPep Team.