ZipTok3D: High-Fidelity 3D Tokenization with Compact Token Prefixes
ZipTok3D claims 32-token reconstruction quality with as little as one token on ShapeNet.
The paper proposes a 3D tokenizer built around compact global-token prefixes rather than spatial latents or fixed global token sets. During training, nested dropout forces shortened prefixes to reconstruct the full object, pushing core geometry into the earliest tokens. Its decoder then iteratively unfolds those prefixes with a shared Transformer block, without a separate sampling stage. The authors report matching a 32-token COD-VAE baseline using one token on ShapeNet and four on TRELLIS. HF Daily Papers' note
The paper proposes a 3D tokenizer built around compact global-token prefixes rather than spatial latents or fixed global token sets. During training, nested dropout forces shortened prefixes to reconstruct the full object, pushing core geometry into the earliest tokens. Its decoder then iteratively unfolds those prefixes with a shared Transformer block, without a separate sampling stage. The authors report matching a 32-token COD-VAE baseline using one token on ShapeNet and four on TRELLIS. HF Daily Papers' note
score 5