The paper introduces ZipTok3D, a 3D tokenizer designed to reconstruct high-fidelity shapes from extremely short token sequences using progressively informative global-token prefixes and iterative decoding. During training, nested dropout randomly truncates the latent sequence so that every retained prefix must still reconstruct the full object, forcing essential geometry into the leading tokens; a parameter-shared Transformer block then recovers detail from any prefix without a separate generative sampling stage. ZipTok3D matches baseline reconstruction quality using 32x fewer tokens on ShapeNet and 8x fewer on TRELLIS.
