Why MyFitnessPal's barcode beats their photo scan — and how we fixed it
MFP shipped photo logging in 2023 and still ranks worse than scanning a barcode. The problem isn't recognition — it's grams. Here's what we did differently.
MyFitnessPal users still scan barcodes 5× more often than they snap photos. After dogfooding both for a month, the reason is obvious: barcode gives you grams baked in. Photo scan guesses, and the guess is bad.
Why MFP's photo scan misses
We tested 20 home meals against MFP's photo logger. Median grams error: 38%. The model recognised "chicken" and "rice" fine — it had no clue if it was 80 g of chicken or 250 g.
What we did
Three changes:
- Plate reference calibration. Volya's prompt anchors estimates to a standard 10" plate (~25 cm). The model is told to convert visual area into grams using palm = 100 g, fist = 200 g, thumb tip = 1 tbsp.
- Editable preview. After scan, every gram is editable. Slider drag auto-rescales macros so the user can correct without losing the breakdown.
- Vegan/diet bias. If profile = vegan and scan shows "chicken-shaped object", Volya asks "tofu or seitan?" before committing.
Median error on the same 20 meals: 15%. Still not perfect — soups and dark-restaurant shots stay hard — but inside the noise of manual logging.
The result
3 taps from "I'm about to eat" to "logged + macros saved". That's the friction unlock that keeps people logging past week 3 (the average MFP dropout point).