# How accurate are AI calorie counts from a photo?

Source: https://nume.co.in/calorie-accuracy/ · Checked on 12 November 2026

Every way of counting calories is an estimate: a photo, the label on a packet, the figure on a menu, a food diary, even a kitchen scale. A photo estimate gets much closer when it comes with a few words about the meal, the amounts, and your own usual portions. Nume is built around all three, and when it isn't sure, it asks or shows you what it assumed.

## Every way of counting is an estimate

Each method has its own error:

| Way of counting | What the evidence says |
|---|---|
| Packet label | US rules allow the true calories to be up to 20% above the figure on the label. |
| Restaurant menu | Right on average across 269 items, but 19% of items had at least 100 kcal more than stated. |
| Food diary | In a review of 37 studies, about 30% of people under-reported, and energy was under-reported by about 15%. |
| Weighing at home | The amounts are exact, but food values are averages. The standard values overstated the calories in almonds by 32%. |
| AI from a photo alone | About 30 to 36% off on average for ChatGPT and Claude in two 2025 studies (Gemini was worse, over 60%), with large portions undercounted most. In a 2026 conference abstract from NIH researchers, four photo apps came in about 250 to 345 kcal low per meal. |

So the useful question is what brings an estimate closer, and whether you can see and correct it.

## What brings a photo estimate closer?

One 2025 study gave an AI model 195 dishes in steps. Adding facts a photo can't show, such as the kind and amount of fat, the sweetener and the kind of meat, brought its error down by about a fifth. Adding the ingredients and their amounts more than halved it. Taking the photo away again made it worse, so the photo and the words work best together.

Another 2025 study found that the bigger the portion, the more the models undercounted it. Saying how much was there matters most on a big plate.

In practice:

- **Add a few words.** What's in it, how it was cooked, how much you ate: "about half the rice".
- **Give amounts in your own units.** A cup, a bowl, two slices, a handful. No scale needed.
- **Use your usual portion.** Your own bowl of pasta tells more than a database average.

**In Nume:** a photo can carry a note, such as "lunch, about half", and Nume reads your words and the photo together. It sizes the whole plate first, then splits it into parts that add up to it, each with a household amount, its weight, energy and protein. A log with words is read next to your own recent similar meals, with the portions you settled on. Type the same words again and you get the same numbers straight away.

## Why do photos miss?

- **Depth.** A photo can't see how deep a bowl is or what's under the top layer.
- **Cooking fat.** Oil, butter and sauces are energy you can't see.
- **Set meals and mixed dishes.** A combo or a platter has more on it than the eye picks out.
- **Shared plates.** A photo of a family order doesn't say how much was yours.

**In Nume:** cooking oil or butter can be its own part of the plate. A named set meal or combo brings in the items it usually has. A shared order is read as the whole order, and Nume asks once how much of it was yours.

## How can a number be checked?

Energy comes from protein, carbs, fat, fibre and alcohol, so an estimate whose energy doesn't match what's in it is wrong somewhere. The same goes for a plate whose parts don't add up to the whole.

**In Nume:** code checks every estimate before it's kept. The energy has to agree with the protein, carbs, fat, fibre and alcohol in it. No part can be denser than pure fat, no part is counted twice, and one log can't hold more than one person's meal. A weaker estimate is marked "≈", logged meals read "about" a number, and planned meals show a range, so you can see how firm each number is.

## Meals from your plan start with numbers

A meal you eat as planned needs no estimate at the table.

**In Nume:** a suggested plate arrives with every part, its amount, energy and protein already worked out, from one stored value per food. The same food has the same numbers on every plate. "Had this" records it as shown, with any part you swapped. Amounts come in hands, bowls, cups and pieces.

## Can I see and change every number?

Look for an app that shows the parts behind each number and lets you change them. A total on its own can't be checked.

**In Nume:** each meal opens to its parts, each with an amount and its energy. Choose how much of the plate you had (all, most, half, a little, or about one and a half times) in one tap. Change a part's amount, remove it, or add something you had, and only the parts you changed are worked out again. What Nume assumed is marked: a share, a time, or the eating time it placed a log in, and when it can't tell which eating time a log belongs to, it asks.

## Is it close enough to lose weight?

What matters for your weight is the pattern over weeks, not one meal. If the scale hasn't moved for a while, [why am I not losing weight?](/not-losing-weight/) goes through the usual reasons, and the [calorie calculator](/tools/calorie-calculator/) gives a starting figure for the day.

## Logging a meal in Nume

Take a photo, type a line, or both. The meal is saved at once and its parts fill in shortly after, each with an amount, energy and protein. Add a few words when you can, correct any part with a tap, and Nume keeps your portions for next time. Meals from your plan arrive already counted. See [how Nume works across a day](/how-it-works/), or [check a plan you already have](/check-my-plan/).

## Common questions

### Can ChatGPT count calories from a picture?

Yes, as an estimate, and the table above shows how far a photo alone can be off. Words and amounts bring it much closer. For how a chat and Nume differ once the meal is logged, see [ChatGPT or Nume](/compare/chatgpt/).

### Do I need a food scale?

No. Amounts in cups, bowls, slices and pieces, plus a few words with the photo, close most of the gap. Weigh if you enjoy it.

### Why do two apps give different numbers for the same meal?

They assume different portions, use different values for the same food, and count cooking fat differently. Saying the amount and how it was cooked removes most of the difference.

### Are restaurant meals harder to count?

They are for any method, since you don't see how the dish was made. Menu figures are a good guide with some dishes running over; see [what to order when you eat out](/eating-out/).

## Sources

- [Rodríguez-Jiménez et al., Image-based dietary energy and macronutrients estimation with ChatGPT-5: cross-source evaluation across escalating context scenarios, Nutrients (2025)](https://pmc.ncbi.nlm.nih.gov/articles/PMC12655113/)
- [Fridolfsson et al., Performance evaluation of 3 large language models for nutritional content estimation from food images, Current Developments in Nutrition (2025)](https://doi.org/10.1016/j.cdnut.2025.107556)
- [Charles et al., Photograph-based AI features in calorie-tracking apps underestimate energy content of meals, NUTRITION 2026 abstract, Current Developments in Nutrition](https://doi.org/10.1016/j.cdnut.2026.108738)
- [American Society for Nutrition: Photo-based calorie-tracking apps may underestimate energy in meals (2026)](https://www.eurekalert.org/news-releases/1136415)
- [Urban et al., Accuracy of stated energy contents of restaurant foods, JAMA (2011)](https://pubmed.ncbi.nlm.nih.gov/21771989/)
- [US Code of Federal Regulations, 21 CFR 101.9(g)(5), nutrition labeling of food](https://www.ecfr.gov/current/title-21/chapter-I/subchapter-B/part-101/subpart-A/section-101.9)
- [Poslusna et al., Misreporting of energy and micronutrient intake estimated by food records and 24 hour recalls, British Journal of Nutrition (2009)](https://doi.org/10.1017/S0007114509990602)
- [Novotny et al., Discrepancy between the Atwater factor predicted and empirically measured energy values of almonds in human diets, American Journal of Clinical Nutrition (2012)](https://pubmed.ncbi.nlm.nih.gov/22760558/)