← back to the accuracy dashboard
Each model's read on the shipped flow (photo + a terse user caption) vs ground truth, for the dishes the models collectively get closest and furthest. Each macro cell shows the value and its signed % error (+ = over-estimate): ≤15% / ≤35% / >35%. Ranked by mean error across calories, fat, carbs & protein over 99 dishes all 4 models scored; mass shown as grams.
Simple, well-separated plates where the photo + caption pin the portions.
Dense, mixed, or visually ambiguous plates where portion size is hard to read.
Nutrition5K camera-C frame 10. Condition macroshot_cam_text_terse. GT from data/prompts.json.