draws a scene as millions of soft blobs, called splats, blended back to front. That order changes with the camera, so web and phone viewers re-sort every splat every frame, and the sort dominates the frame.
The method groups splats into small boxes in a nested tree that halves space along each axis, and stores each box in depth order along a fixed axis. Each frame sorts only visible boxes, and one product of two short vectors per box decides whether to read it forwards or backwards. A short training pass adapts an existing scene.
Quality stays close, losing 0.13 decibels of peak signal-to-noise ratio, a sharpness score, and renders 4.5 times faster where the processor sorts.