8 Inputs → 58 Body Parameters

Trained a 2-layer × 256 MLP that predicts Anny's 58 body shape parameters from just 8 survey inputs (height, weight, gender, body type, frame size, waist, cup size, ethnicity). The model size is about 85 KB, and training time on a laptop is about 120 minutes. The architecture is ordinary, but the loss design is the key.

On a held-out synthetic test set, the errors are as follows.

• Female: height 0.3/0.8 cm, mass 0.4/1.0 kg, bust/waist/hips 2.7/4.0/3.3 cm
• Male: height 0.3/0.8 cm, mass 0.5/1.2 kg, bust/waist/hips 4.9/4.3/3.3 cm

Since mass is not a direct output of the 58 parameters but a result of volume, the path MLP output → Anny blendshapes → vertices → signed tetrahedra volume → predicted mass/height was put directly into the loss and backpropagated through. Because Anny uses linear blendshapes and differentiable volume, this is handled by PyTorch autograd without any custom backward pass.

For comparison, Bartol et al. (2022)'s height+weight linear regression had a BWH MAE of about 7 cm on the same set, and the SAM 3D Body → MHR → Anny + tuning photo pipeline achieved 5-8 cm against real people without SMPL. This shows that survey inputs can carry more signal than single-image HMR.