Nishi 3D Gaussian Splatting — the 2025 frontier rep, on our OWN stack

The current SOTA for photoreal avatars (GaussianAvatars, HumanSplat, VRGaussianAvatar) renders a cloud of oriented 3D Gaussians. This is that pipeline, 100% sovereign: our inside-out anatomical base → a surface-nets mesh → anisotropic surface Gaussians (surfels aligned to the surface normal, projected to 2D ellipses) → depth-sorted, alpha-composited by our own integer splatter — then photo-coloured from our Z-Image reference. No WebGL, no libraries, no trained weights (only a generator needs training; the renderer is fully ours). Honest tier: hand-placed surfels on the mesh, not yet gradient-fit to images (the generator rung); a stylized base, real frontier rendering.

loading…

The generator — fitting Gaussians to the real photo

The other half of 3D Gaussian Splatting is the optimizer that fits the Gaussians to a target image — also sovereign + integer. A splatted pixel is linear in the Gaussian colours, so the image-loss gradient is analytic (no autograd, no float). Here the target is the actual reference photograph: 12,288 Gaussians start grey and gradient descent fits them to reproduce it. Left: the real photo. Middle: the grey start. Right: the fitted Gaussian splat — a recognizable reproduction (image loss down to 6%), rendered by our own integer splatter. Honest boundary: this is a single-view fit (a 3D-Gaussian representation of one 2D photo) — it reproduces the face from the front; a full 3D-consistent avatar from one photo needs single-image-3D reconstruction (a trained net), the next rung. The renderer + optimizer are entirely ours.