MorselFlow

What one benchmark found

Why two food-photo algorithms struggled with simple foods

Researchers tested two public, recipe-trained algorithms on food photos from 95 U.S. adults. The algorithms predicted ingredients less accurately for single-food images than for images containing several foods.

The SNAPMe database links 3,311 unique photos to 275 detailed food records. Participants photographed everything they consumed and completed a separate dietary record for up to three study days, giving researchers a reference for checking ingredient predictions.

SNAPMe paired food photos with records completed by the same participants.

Algorithm inputFood photographs
ReferenceParticipant food records
The study recruited generally healthy U.S. adults ages 18–65 who recorded food for up to three days. The 3,311 photos are not 3,311 independent people or meals. Source: SNAPMe: Surveying Nutrient Assessment with Photographs of Meals

The study measured ingredient prediction, not whether every AI can name a dish.

Researchers evaluated FB Inverse Cooking and Im2Recipe, two public algorithms trained largely on recipe data. For Im2Recipe, the mean F1 score was 0.08 for images matched to one food code, 0.15 for images with two or three food codes, and 0.18 for images with four or more food codes. A higher F1 score indicates a better balance of correct ingredient matches and missed or incorrect predictions.

The paper also reports difficulty with beverages. The authors suggest that recipe-focused training data may contain less useful coverage of basic foods and drinks, but the study did not prove that every modern image model will make the same errors.

A visually simple item still deserves a quick check.

When an app suggests a food from a photo, check details the image may not reveal: whether a clear drink is water or a sweetened beverage, whether a plain-looking item contains a filling, and whether a sauce sits outside the frame.

These checks take seconds because you already know what you ate. The image model can start the record; your knowledge resolves information that pixels alone do not contain.

MorselFlow lets you resolve what the photo leaves ambiguous.

MorselFlow shows suggested foods and portion estimates before saving. Check clear drinks, plain single foods, packaged products, and anything outside the frame first.

If the suggestion is wrong or incomplete, change it before adding the meal to your timeline. SNAPMe did not evaluate MorselFlow, so the paper cannot provide an accuracy rate for the app’s current model.

The figures in this article come from population guidance or research. They are not personal diagnosis, treatment, or nutritional advice. Food, portion, and nutrition suggestions are estimates and may differ from the actual meal.