neural-implicit-fields
A coordinate-based MLP f(x, y, z) → signed distance was fit to a target
shape in PyTorch. The surface below is the zero level set of that single learned
field, extracted by marching cubes at three different resolutions — the triangle
count is a choice made at extraction time, not a property of the model.
The field is optimised with signed-distance regression plus an Eikonal regulariser (‖∇f‖ = 1), starting from a geometric sphere initialisation. The centre image is the learned field rendered by differentiable sphere tracing with autograd normals.
A slice of the learned signed-distance field with its zero level set, and the same field meshed at 24³, 48³, and 128³. The interactive model above shows exactly these extractions.
The same kind of field learned from depth and silhouette images alone: the loss is computed on rendered pixels and back-propagated through the unrolled sphere march into the network weights — no ground-truth distances involved.
