Each of the 516 objects is a real 3D model from Objaverse, encoded by a triplane variational autoencoder into a latent of 3 × 32 × 32 × 8 numbers. A 1.31 M-parameter decoder turns any latent into a signed-distance and colour field, evaluated on a dense grid on your GPU (WebGPU compute or WebGL shaders) and sphere-traced every frame. Walking blends latents, so every in-between frame is a shape the network invents.
loading…
Partners are neighbours by decoded shape overlap, not by latent distance: on 32 measured pairs the in-between shapes keep 97 % of their volume with 3 % floating debris, against 75 % and 17 % for latent-nearest pairs. The bottom-up sweep halves the debris again. The map is a t-SNE of the latents, so distances on it are approximate.
space pause · → next · ← back · 1–6 walks · s save png · / find · m map · r spin · h HD · i about · d debug