8 inputs → 58 body params: putting a body-model forward pass inside the training loss [P]
Training loss includes body model forward pass for mass and height accuracy
A new approach to body-shape prediction uses a tiny MLP (multilayer perceptron) with just 2 layers and 256 units, totaling roughly 85KB of parameters. The model takes 8 simple questionnaire inputs—height, weight, gender, body shape, build, belly, cup size, and ancestry—and outputs 58 parameters for the Anny body model. Training completes in about 120 minutes on a standard laptop.
The critical innovation lies in the loss function. Instead of predicting body parameters independently, the training loss incorporates the full body model forward pass. The MLP outputs go through Anny's linear blendshapes to produce a mesh, then compute volume via signed tetrahedra, mass using a two-component density model, height from vertex positions, and waist circumference via ISO 8559-1 plane-sweep algorithms. All these differentiable operations allow gradients to flow back through the entire chain, enforcing physical consistency. The results are striking: height MAE of 0.3 cm, mass MAE of 0.4-0.5 kg, and bust/waist/hips MAE of 2.7-4.9 cm. For comparison, ridge regression baseline achieved 3.9 kg mass MAE, showing a 10x improvement from the physics-aware loss alone.
- 85KB MLP achieves 0.3 cm height MAE and 0.4 kg mass MAE, 10x better than ridge regression
- Physics-aware loss includes forward pass through body model for volume, mass, and height
- Training takes ~120 minutes on a laptop; relies on custom measurement library (clad-body, Apache 2.0)
Why It Matters
Enables accurate body-shape prediction from simple inputs, useful for virtual try-ons, health, and ergonomics.