How accurate is AI calorie counting from a photo?
AI can turn a photo of your meal into a calorie estimate in seconds. Here is how it works, why the numbers can be off, what studies show and how to use them wisely.
Photo calorie counters, including PlateCalories, are handy, but how close are their numbers to what you actually ate? In short: a photo gives a useful estimate, not a measurement. In three 2025 studies, general AI models working from a photo alone had average calorie errors from about 30% to over 70%, depending on the model.
How AI estimates calories from a photo
There are three tasks, done by separate tools or by a single AI model:
- Recognizing foods: identifying each food and drink, such as rice, grilled chicken and a glass of milk.
- Estimating portions: judging how much of each item there is from visual clues, such as its size compared with the plate or cutlery.
- Converting to energy: turning each amount into calories using typical values for that food, then adding up the total.
Errors can creep in at each step and add up. PlateCalories gives all three tasks at once to a single general-purpose AI model, which lists each food with its estimated portion and calories; PlateCalories then adds up these calories to give the total. The model does not weigh food or look up a nutrition database, so its values are typical, not measured.
Which AI models PlateCalories uses
PlateCalories relies mainly on open models, which their makers publish so that others can download and run them: Gemma 4 from Google, GPT-OSS from OpenAI and Llama 4 Scout from Meta. Which service answers depends on the country you connect from, as detected by Cloudflare. PlateCalories uses the free tiers of other AI services first and keeps Cloudflare Workers AI for last. Each request goes to the first service listed below for your region; if that service is busy or unavailable, the next one takes over:
- Europe (the European Union and the rest of the European Economic Area, the UK, Switzerland, Gibraltar, the Channel Islands, the Isle of Man, the French overseas departments and Europe's micro-states, plus any visitor whose country cannot be detected and visitors from Iran, North Korea, Cuba and the Ukrainian regions of Crimea, Sevastopol, Donetsk, Luhansk, Zaporizhzhia and Kherson): a meal photo, with any description you add, is analyzed on Cloudflare Workers AI by Gemma 4 or, as a backup, Llama 4 Scout. Descriptions without a photo, typed or spoken, go first to GPT-OSS on Groq (USA), then to Gemma 4 on Cloudflare Workers AI. Google, Z.ai and OpenRouter are never used for these analyses.
- United States (Puerto Rico and the other US territories included): a meal photo, with any description you add, goes first to Gemma 4 through the free tier of Google's Gemini API, then to Gemma 4 through OpenRouter (whose free Gemma model is run by Google AI Studio), then to Cloudflare Workers AI, which uses Gemma 4 and, as the last backup, Llama 4 Scout. Descriptions without a photo go first to Gemma 4 through Google's Gemini API, then to GPT-OSS on Groq, then to Gemma 4 on Cloudflare Workers AI. Z.ai is never used.
- Everywhere else (for example Canada, Australia, Africa, Asia and Latin America): the same order as in the United States, with Z.ai (Zhipu AI's international platform, based in Singapore) as an extra backup right after Google.
Different models, and even the same model on a second try, can give somewhat different figures, so the same photo will not always get exactly the same result. PlateCalories does not store your photos, voice recordings, descriptions or results, but each AI service has its own rules. For example, Google may use what is sent to the free tier of its Gemini API to improve its products, and human reviewers may read it; requests sent through OpenRouter to its free Gemma model, which Google AI Studio runs, may be kept for up to 55 days. Our privacy policy gives the full details.
The main sources of error
Portion size and depth
A photo is flat. It shows the surface of the food, not its height or what lies underneath, so a thick slice of lasagna or a deep bowl of soup can look smaller than it is. In a 2025 study, some of the largest errors were for dishes served in bowls, possibly because depth is hard to perceive. In the Nutrition5k research project, adding data from a depth sensor cut a model's average calorie error from about 26% to 17–19%. PlateCalories uses an ordinary photo, so it helps to mention a deep bowl or a big serving.
Hidden fats and sugars
Oil, butter, sugar and creamy sauces are often invisible or mixed into food, yet rich in energy. According to USDA FoodData Central, olive oil has 884 kcal per 100 g, or about 119 kcal per tablespoon (13.5 g): two tablespoons in a stir-fry add about 240 kcal that the camera may miss. Our guide to hidden calories gives more examples.
Mixed dishes
In stews, curries, sandwiches and burritos, ingredients are layered or blended, so the AI has to guess what is inside and how much. In another 2025 study, ChatGPT-4o estimated a large serving of lentil curry at 255 g; it actually weighed 480 g.
Lighting, angle and framing
Dim light, strong shadows, filters, extreme angles or a plate cut off at the edge make foods harder to recognize and portions harder to judge. Our food photo tips explain simple fixes.
Unusual or regional dishes
AI models learn from examples, and some cuisines and family recipes may be rare in their training data. In a 2024 study of food-tracking apps, AI image recognition could not identify the separate ingredients of beef pho (a Vietnamese noodle soup) or pearl milk tea. The apps' energy estimates were up to 49% too high for the pho and up to 76% too low for the tea. The authors called for training AI on more mixed dishes and culturally diverse foods.
What research says
- Close to human estimates: a systematic review of 52 studies (2010–2023) found average calorie errors from 0.1% to 38.3% for AI methods that analyze food images, with smaller errors for simple, single foods. The authors judged this in line with human estimates, but said the tools needed more development before being used on their own.
- Foods are easier than amounts: on 114 meal photos, ChatGPT-4 named foods well (93% of the foods it listed were really there) but underestimated the meal's weight in 76% of the photos, mainly for medium and large meals (O'Hara et al., 2025).
- Bigger portions, bigger errors: on 52 food photos, ChatGPT-4o and Claude 3.5 Sonnet had an average energy error of about 36%, and Gemini 1.5 Pro about 64%; all three underestimated more as portions grew. The authors judged the first two comparable to traditional self-reported methods such as food diaries, but not yet precise enough for clinical or sports nutrition (Fridolfsson et al., 2025).
- The model matters: in a 2025 benchmark of seven general AI models on 505 photos of cafeteria meals, with no information on food weight, average calorie errors ranged from about 33% to 73%. Llama 4 Scout, which PlateCalories keeps as the last backup for meal photos, had the highest error. Gemma 3 27B, an earlier-generation Gemma model from Google, had the second-highest error, about 55% (Mu et al., 2025).
- Details help: with ChatGPT-5 on 195 dishes, the average error was about 123 kcal per dish (31%) from the photo alone, 92 kcal (24%) with details such as the type and amount of fat and sweetener, and 53 kcal (14%) with a full list of ingredients and amounts. The same list without the photo gave a larger error, about 67 kcal or 18% (Rodríguez-Jiménez et al., 2025).
Each study tested specific models in specific conditions. None of them tested Gemma 4, the model PlateCalories tries first for meal photos, and PlateCalories itself has not been evaluated in a published study, so its accuracy may be better or worse.
How to use AI calorie estimates wisely
- Treat each result as approximate. A single number can be well off, especially for large meals and mixed dishes.
- Follow trends over days and weeks. Over many meals, random highs and lows partly even out, so patterns tell you more than any single result. A consistent bias does not even out, though.
- Allow for big portions. Two of the studies above found that AI underestimated large servings, so if you eat generously, your totals may run low.
- Check foods you eat often. Weigh your usual breakfast once, check the label or a food database, and compare with the AI estimate. No scale? See our guide to portion sizes without a scale.
How to add details in PlateCalories
In the PlateCalories calorie counter, you can add details to your photo by typing a description (up to 500 characters) or by speaking for up to 60 seconds. If your browser offers voice typing, its own speech recognition service turns your words into text (Google's in Chrome, Microsoft's in Edge, Apple's in Safari, unless the browser does this on your device), and only that text reaches PlateCalories. If it does not (for example in the built-in browser of the WhatsApp, Instagram or Facebook apps), or if your browser's main language is not English, Spanish or French, your voice is recorded instead, so that the language you speak can be recognized. The recording, like any audio file you choose, is sent once to be transcribed by Whisper, an open speech-recognition model from OpenAI, on Groq (USA) or, as a backup, on Cloudflare Workers AI. The AI then uses your words to correct food names, amounts and hidden ingredients. For example:
- cooking fat: “fried in about 2 tablespoons of olive oil”;
- known amounts: “200 g of cooked rice”;
- drinks and extras: “coffee with 2 teaspoons of sugar”, “regular cola, not sugar-free”;
- fillings: “burrito with beans, cheese and sour cream”.
Check what was understood: dictated words appear in the description box, where you can correct them, and the transcript of a recording is shown with the result. The result lists each item with its estimated portion and calories. If something is missing or looks wrong, add a correction and analyze again (daily usage limits apply). You can also describe a meal by voice or text, without a photo. The estimate then rests on your words alone, so include amounts when you know them.
When precise tracking matters
Day to day, a rough estimate is often enough. Some situations call for more precision and professional support:
- diabetes, especially if you count carbohydrates to adjust insulin (PlateCalories estimates calories only, not carbohydrates);
- kidney disease, other medical conditions or a diet prescribed by a doctor;
- pregnancy or breastfeeding;
- a current or past eating disorder.
In these cases, ask a doctor or registered dietitian which method suits you, such as weighing your food. Never use an AI estimate to decide a medication dose, such as insulin. If tracking calories makes you anxious or leads you to eat very little, take a break and talk to a health professional. PlateCalories is for adults (18+) and is not medical advice.
Sources
- Shonkoff et al., Annals of Medicine (2023): systematic review
- O'Hara et al., Nutrients (2025)
- Fridolfsson et al., Current Developments in Nutrition (2025)
- Rodríguez-Jiménez et al., Nutrients (2025)
- Mu et al., ACM BCB (2025)
- Li et al., Nutrients (2024)
- Thames et al., CVPR (2021): Nutrition5k
- USDA FoodData Central: olive oil (FDC ID 171413)