OmniFood-Bench - Portion Size (Homemade Food): leaderboard
OmniFood-Bench asks vision-language models to estimate the weight in grams of the components of a food photograph. This board covers images of home-cooked dishes and is scored as mean absolute percentage error, so lower is better and the value is unbounded above.
Metric: Portion-size MAPE (%; mean absolute percentage error of the predicted weight in grams of the visible components of Homemade Food images, lower is better and unbounded above; images of the MM-Food-100K subset; zero-shot, one standardized prompt per task; the open-weight models see all 1,208 samples and the proprietary models a 496-sample subset). Source: arxiv.org. Saturation forecast: Around May 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4 | 51.6 | #400 |
| 2 | GPT-5.1 | 60.54 | #229 |
| 3 | Gemini 3 Flash | 74.99 | #164 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=omnifood-bench-portion-size-homemade-food · How It Works · Data refreshed daily, snapshot 2026-11-13.