Researchers have introduced ZipTok3D, a novel 3D tokenizer designed to achieve high-fidelity reconstruction from extremely short token sequences. The method addresses a key limitation in existing 3D tokenizers, which often suffer sharp reconstruction degradation when compressed to very low token budgets.
Traditional approaches organize latent representations either over spatial regions or as fixed-size sets of global tokens. ZipTok3D instead organizes object geometry into progressively informative global-token prefixes, which are then unfolded through iterative decoding. This allows the model to prioritize essential geometric information in the leading tokens.
The training process employs nested dropout, which randomly truncates the latent sequence after encoding and requires each retained prefix to reconstruct the complete object. This encourages the model to place the most critical information in the earliest tokens. During inference, a parameter-shared Transformer block is applied iteratively to recover fine-grained geometry from each prefix, eliminating the need for a separate generative sampling stage.
Experimental results show that ZipTok3D achieves reconstruction quality comparable to the 32-token COD-VAE baseline using only one token on ShapeNet and four tokens on TRELLIS. This represents a 32x and 8x reduction in token sequence length, respectively, while maintaining high fidelity.