Thin wrapper around the Inria differentiable 3D Gaussian rasterizer. Renders a batch of Gaussians into the 4D virtual camera defined by ``ren

4x transposed-conv stack that lifts ``em`` (B, 512, 4, 16) back to (B, C, 64, 256).

Negative Pearson loss along the temporal dimension. Inputs are expected to be 3-D tensors ``(M, N, T)`` — typically ``M`` = batch, ``N`` = nu

Spectral loss for heart-rate estimation. Computes the magnitude spectrum of a Hanning-windowed BVP signal at the candidate BPM bins ``[low_bo

ContrastPhys-style spatio-temporal consistency on ``v_d``. ``v_d`` shape: ``(B, N=64, T=256, 3)``. The ``N`` rows are split into ``K`` regions