How accurate is PixNutri?

Short answer: PixNutri gives useful estimates, not exact measurements. A photo can’t weigh your food or see the oil it was cooked in, so any single scan can be noticeably off. That’s why every result in PixNutri can be corrected in one tap.

We would rather be honest about this than promise “precise” numbers.

What independent research says

AI models reading food photos. A 2025 study in Current Developments in Nutrition tested three general-purpose AI models on 52 standardised food photos, with food weighed and analysed against a nutrition database as the reference. For energy (calories):

Model tested	Average error in calories (MAPE)
ChatGPT-4o	35.8%
Claude 3.5 Sonnet	35.8%
Gemini 1.5 Pro	64.2%

The authors concluded that the best models reached accuracy comparable with traditional self-reported food diaries, but that they systematically underestimated large portions and are not yet suitable where precise quantification is critical, such as clinical nutrition.

Food diary apps in general. A 2021 meta-analysis in Advances in Nutrition found that diet-tracking apps underestimated energy intake by about 200 kcal per day on average compared with reference methods (pooled effect −202 kcal/d). The gap was much smaller when the app and the reference used the same food database.

In other words, error is normal in every kind of food tracking, whether you type, scan or photograph. What matters is knowing where it comes from and correcting the big mistakes.

What we know, and don’t yet know, about PixNutri

• PixNutri uses Gemini 3.1 Flash-Lite, a newer model generation than the Gemini 1.5 Pro tested in the study above. That study therefore doesn’t tell us how PixNutri performs.
• We have not yet published our own benchmark. We plan to test PixNutri against weighed meals across different cuisines and portion sizes and publish the results, including where it does badly. Until then we won’t put an accuracy percentage in our marketing.
• Barcode scans don’t depend on AI recognition: they use the nutrition facts printed on the product, as recorded in Open Food Facts, scaled to the amount you enter. Their accuracy depends on that database entry and on the amount you type.

Where estimates usually go wrong

1. Portion size. The biggest source of error. A photo has no scale, and larger portions are the most likely to be underestimated.
2. Hidden fat and sugar. Cooking oil, butter, dressings and sauces add a lot of calories and are often invisible.
3. Mixed and layered dishes. Rice under curry, fillings inside wraps, toppings on pizza.
4. Look-alike foods. Full-fat vs light yoghurt, sugar-free vs regular drinks, white vs whole-grain bread.
5. Photo quality. Dark, blurry or very close photos give the model less to work with.

How to get better results

• Photograph from above in good light, with the whole plate in view.
• Check the portion and use the ½ to 2× buttons when the serving looks bigger or smaller than estimated.
• Add what the photo can’t see. If something was fried or had a creamy sauce, raise the calories or fat.
• Use the barcode scanner for packaged food whenever you can.
• Look at trends, not single meals. Small errors in individual scans matter less than your weekly pattern.

If you need exact numbers for a medical condition, such as diabetes, kidney disease, an eating disorder or pregnancy, work with a doctor or registered dietitian and weigh your food.

Want the technical details? Read our methodology.