AI calorie estimates from photos can be useful starting points, but their accuracy depends on identifying the food, judging its portion and knowing how it was prepared. A photograph cannot directly weigh a serving or reveal every ingredient. There is no single accuracy percentage that applies to all meals and models.
- A photo can help identify visible foods. It cannot directly measure their weight or reveal hidden ingredients.
- Even in controlled tests, image-based AI estimates have missed larger portions.
- Correcting the portion and preparation often matters more than debating the last few calories.
- In Burnline, an AI result is reviewed before it becomes a diary entry; manual logging remains available.
What the camera knows
The image can show recognizable foods, their visible proportions and sometimes a useful size reference such as a fork or plate. That is enough to suggest a list of foods and a starting quantity. The calorie estimate then needs a food-composition value and a portion weight. Those are separate choices, not pixels inside the photograph.
USDA FoodData Central, for example, reports nutrient values on a 100-gram basis and publishes gram weights for some household portions. The arithmetic is straightforward once the food and portion are known; the difficult part is deciding whether the pictured serving matches either one. See the USDA documentation and our note on where calorie counts come from.
What stays outside the frame
Two unknowns matter especially. Portion size is a three-dimensional quantity inferred from a two-dimensional view. The same mound of rice can look different with a different plate, angle or depth. Preparation is often invisible: oil absorbed during cooking, butter under vegetables, sugar in a drink, or a sauce mixed into a dish. The image may also hide a second serving or a food eaten later.
This is not merely a hypothetical AI failure. In the Nutris-Phone study, a trained nutrition scientist evaluated photographed meals against weighed food records. The researchers described add-ons and unobservable foods as obstacles to accurate estimates, even with a human expert in the process. Read the original study in Public Health Nutrition.
A photo starts the estimate.
You fill in the gaps.
Recognizable food is a starting point.
- Food type → check itDoes the suggested list match what you ate?
- Portion weight → add itA single view cannot directly weigh the serving.
- Oil & sauce → describe themPreparation and mixed-in ingredients may be invisible.
How accurate is AI, then?
There is no single accuracy percentage for “AI from a photo.” It depends on the image, the meal, the reference used to judge it and the particular model. In one 2025 validation study, researchers tested three multimodal models on 52 standardized food photographs. They weighed the foods and used a nutrition database for comparison. All three models increasingly underestimated larger portions, despite being given visible plates and cutlery as size references. The study is available through PubMed.
Those results are a reason for caution, not an accuracy score for Burnline or for every model available today. A controlled photograph of a known meal is also different from a restaurant dish with an unknown recipe. A plausible-looking number should not be mistaken for a measurement.
The most useful question is not “Did the AI give me an exact number?” It is “Which assumption would change this estimate most if it is wrong?”
A better way to use the draft
Start with the visible foods. Is anything missing or mislabeled? Then check the quantity: is that one serving, two servings, or only part of what you ate? If you know the weight, use it. For a packaged food, the label or a barcode may be a better starting point than a photo.
Add what the camera cannot know. “Pan-fried with about a tablespoon of olive oil” or “dressing on the side, half used” is more informative than asking the model to be more confident. For a homemade dish, calculate the portion from its ingredients and cooked yield when those details are available. If you do not know those details, keeping the result approximate is more honest than pretending the uncertainty has disappeared.
Finally, look at the result before saving it. An estimate can help reduce typing; it should not silently become a fact in your diary. Nor should one uncertain meal become a reason to restrict food or exercise to “correct” a number.
Where Burnline draws the line
Burnline offers a photo or text estimate as an optional path for a meal that is awkward to find in a list. It asks permission before sending the selected input to its AI processors. The result is shown for review and editing before saving; choosing not to use AI leaves the manual and built-in food paths available. The Privacy Policy names the processors and explains what is sent.
The boundary matters: a photograph is an observation, the model's answer is an inference, and the saved entry is your record. We should not make those three things look identical.
This article explains dietary-assessment research and a tracking interface; it is not medical or individualized nutrition advice. AI estimates are not appropriate as a substitute for a clinician or dietitian where precise intake is medically important.
Sources
- Fridolfsson J, Sjöberg E, Thiwång M, Pettersson S. Performance evaluation of 3 large language models for nutritional content estimation from food images. Current Developments in Nutrition. 2025.
- Prinz N, Bohn B, Kern A, et al. Feasibility and relative validity of a digital photo-based dietary assessment: results from the Nutris-Phone study. Public Health Nutrition. 2019.
- USDA Agricultural Research Service. FoodData Central: Foundation Foods documentation. Updated 2024.
