The SNAPMe database contains 3,311 food photos linked with 275 food records from 95 participants. Public algorithms tested in the study performed poorly especially for single-ingredient foods and beverages.
The number at a glance
More photographs do not remove information that was never visible.
Simple to us, ambiguous to a model
A simple item may provide fewer visual clues.
Water and an unsweetened drink, a hidden sauce, or similarly shaped single foods can be difficult to distinguish from color and outline alone.
A composite dish may offer many recognizable features, while a simple item can leave the model with less evidence.
Design around uncertainty
Editable is more trustworthy than pretending to be exact.
An AI suggestion is a hypothesis based on the image. A product name, amount, or drink outside the frame is often known more reliably by the person.
Treating uncertainty honestly and allowing foods and portions to be corrected is not a weakness. It is a necessary part of the feature.
What to keep
Use the AI suggestion as an editable draft, not as the answer.
How MorselFlow fits
MorselFlow begins AI estimation with a review screen.
It shows suggested foods and portion estimates, and lets you change them before saving.
Items outside the image and difficult drinks are expected to need human addition or correction.
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.