2024 → 2026 準確率曲線,一圖看懂

前十名的識別準確率兩年內提升 31 個百分點。分量準確率提升不到其一半。

---

上方標題與摘要已翻譯;正文目前為英文,我們正在進行在地化。

We have run this benchmark, in some form, since Q1 2024. That gives us a two-year window, enough to see where the field is actually moving versus where it is just marketing.

Identification: solved-ish

Across the top 10, mean identification accuracy went from 53.1% (Q1 2024) to 84.3% (Q1 2026), a 31-point gain. Most of that came from two model generations of vision-language progress, not from food-specific innovation.

The takeaway: if your app still can’t tell pad thai from yakisoba, it’s not a food AI problem anymore, it’s a “this team didn’t upgrade their model” problem.

Portion: still hard

Mean portion error dropped less than half as fast, from ±18.4% to ±7.9% in the same window. And the distribution is bimodal. The top three apps cluster between ±1 and ±5%. The rest of the field is still above ±7%.

The leader, Welling, is roughly five years ahead of the bottom of the chart on this single metric.

Why the gap will widen before it closes

Portion estimation rewards proprietary data and depth fusion. Both are expensive to build and hard to copy from a research paper. Identification rewards using the best available foundation model, which everyone can do.

Expect identification to keep converging at the top of the leaderboard, and portion accuracy to keep separating the leaders from the field.