Neural — tiny MLP SDF trained in the browser

← HomeNeural · learn the field

Experiment. A small MLP (1793 weights) is trained in plain JS to fit the reference SDF. Once trained, the weights bake into a GLSL shader and we raymarch the network. Neural nets aren't Lipschitz — use the Step multiplier to shrink each march step and avoid overshoot.

Analytic (truth) — drag to orbit both

Neural · stale

Samples: 4000LR: 0.050Batches / frame: 8Step multiplier: 0.50

iter

loss

—

surface RMS

—

weights

1793

shader chars

—

Arch: encoded xyz (21) → tanh(32) → tanh(32) → 1. Surface RMS is measured on samples near the true surface — the number you actually want small. If the baked render is blobby, drop the step multiplier.