An 85 KB MLP predicts 58 Anny body-shape parameters from 8 questionnaire inputs using a differentiable physics loss, beating photo pipelines on circumference accuracy with CPU-only inference.
Key Takeaways
The core trick: MLP outputs feed through Anny’s forward pass (linear blendshapes + signed-tetrahedra volume summation), making mass error gradients flow back jointly through all 58 volume-related params – something Ridge regression cannot do.
Accuracy: 0.3 cm height MAE, 0.3-0.5 kg mass MAE, 3-4 cm BWH MAE – better than the author’s own photo pipeline (5-8 cm) and Bartol’s h+w linear regression (7 cm bias-corrected).
Model footprint: two hidden layers of 256 ReLU units, ~85 KB weights per gender, trains on a laptop in 60 min, runs in milliseconds on CPU at inference.
Anny’s anthropometry module had a density bug – it used a flat 980 kg/m3; fix applies the Navy body-fat formula to derive per-person density, closing a 2-3 kg systematic mass error on lean and soft bodies.
Theoretical questionnaire ceiling is ~1 cm waist MAE: roughly 50 continuous blendshape params that have no multiple-choice mapping set a hard floor regardless of how many questions are added.
Hacker News Comment Review
The discrete input space (8 questions, finite h/w buckets) means the full combination set is small enough to precompute offline, reducing live inference to a table lookup and making the millisecond runtime moot.
Notable Comments
@rgovostes: points out the input space is finite enough to precompute ~10M combinations, making results “basically instantly” retrievable without any model at runtime.
