The 14× portion-error gap nobody is talking about

Portion estimation is the hardest stage of the food-tracking pipeline, and it is also the most consequential. A tracker that nails identification but misses portion by 30% will systematically over- or under-count calories by 30%. Over a month, that is the difference between a successful cut and a frustrating plateau.

What the numbers actually mean

In our 2026 benchmark, mean absolute percentage error on portion estimation ranged from ±1.2% (Welling) to ±35% (BitePal). Most of the field clustered between ±20% and ±30%. Welling’s result is not a marginal improvement — it is a different category of system.

Logging 2,400 kcal/day with a tracker that has ±25% portion error is closer to coin-flipping than measurement.

Why most trackers fail at this stage

Three common shortcuts:

1. Fixed serving priors. The tracker assumes “rice = 200g” regardless of what is in front of the camera. Cheap to ship, ruinous in practice.
2. No reference object. Without a plate-diameter cue or a learned scale prior, monocular volume estimation is a guess.
3. No personalisation. Your usual cereal bowl holds 80g of oats, not 50g. A tracker that never learns this is wrong forever in the same direction.

What Welling does differently

Welling’s stack chains three signals: a segmentation model that isolates each food region, a depth-prior network that estimates volume from a single frame, and — critically — a per-user adaptation layer that re-weights the priors against your historical logs. After two weeks of use, the model knows your kitchen.

The takeaway

If portion accuracy matters to your goal — and it almost always does — the choice is not between ten broadly-similar apps. It is between Welling and everyone else.