ZipTok3D: High-Fidelity 3D Tokenization with Compact Token Prefixes
Mingda Lin, Weijie Wang, Zeyu Zhang, Bowen Cui, Yefei He, Haoyu Zhao, Yuanyu He, Donny Y. Chen and 2 others
Zhejiang University · Monash University · University of Adelaide
Published September 1, 2026 · on the feed September 3, 2026 · ▲ 23 · arXiv 2609.01740
The abstract
Compact token sequences are essential for efficient 3D generation. However, existing 3D tokenizers typically organize latent representations either over spatial regions or as fixed-size sets of global tokens, both suffering sharp reconstruction degradation when compressed to extremely low token budgets. In this paper, we present ZipTok3D, a 3D tokenizer designed for high-fidelity reconstruction from extremely short token sequences. Its key idea is to organize object geometry into progressively informative global-token prefixes and unfold these compact representations through iterative decoding. Specifically, nested dropout randomly truncates the latent sequence after encoding during training and requires each retained prefix to reconstruct the complete object, thereby prioritizing essential geometric information in the leading tokens. The decoder then repeatedly applies a parameter-shared Transformer block to recover fine-grained geometry from each prefix without a separate generative sampling stage. With the same token dimension, ZipTok3D achieves reconstruction quality comparable to the 32-token COD-VAE baseline using only one token on ShapeNet and four on TRELLIS, yielding 32times and 8times shorter token sequences, respectively.
The Brief
2:06A short podcast summary of the paper. Written and voiced by AI.
Transcript
ZipTok3D. High-Fidelity 3D Tokenization with Compact Token Prefixes. By Mingda Lin and nine others, from Zhejiang University, Monash University, and one other institution. Published on the first of September. No code repository is linked to it.
Efficient 3D generation depends on compact token sequences: the fewer tokens an object takes, the cheaper it is to generate. Existing 3D tokenizers usually organise their latent representations in one of two ways, either over spatial regions or as fixed-size sets of global tokens. Both suffer sharp reconstruction degradation when compressed down to extremely low token budgets.
ZipTok3D is designed for high-fidelity reconstruction from extremely short token sequences. Its key idea is to organise an object's geometry into global-token prefixes that become progressively more informative, and then to unfold those compact representations through iterative decoding. The first tokens carry what matters most, and each further token adds to it.
Training uses nested dropout. After encoding, the latent sequence is truncated at random, and whatever prefix is retained must still reconstruct the complete object. Since any prefix might be all the decoder receives, the essential geometric information is pushed into the leading tokens. Decoding then applies one parameter-shared Transformer block repeatedly, recovering fine-grained geometry from each prefix. There is no separate generative sampling stage.
The comparison holds the token dimension fixed. Against the COD-VAE baseline, which uses thirty-two tokens, ZipTok3D reaches comparable reconstruction quality using only one token on ShapeNet, and four tokens on TRELLIS. That makes its token sequences thirty-two times shorter on the first dataset, and eight times shorter on the second.
Ordering the tokens by importance changes what truncation costs. Instead of choosing between regions of space or a fixed global set, the representation is arranged so that a shorter prefix still describes the whole object, with the remaining tokens adding geometric detail. On ShapeNet, a single token is enough to stand level with a baseline that needs thirty-two.
Paper Quest · 7 chapters · about 39 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Short sequences make 3D generation cheap. Short sequences also wreck the shapes. Here is the wall.
You are building the step that turns a 3D object into a short sequence of tokens for a generator. The fewer tokens per object, the cheaper generation becomes.
The road ahead
- 1 · The problem ~6 min A 3D generator works on tokens, and squeezing a whole object into very few tokens makes today's tokenizers lose the shape.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Train every random-length prefix to rebuild the whole object, so key geometry moves to the front, then decode by repeating one shared block.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Encode the shape into global tokens, cut them at a random length, rebuild the full object from the prefix, and decode with a repeated shared block.
- 6 · The experiment ~6 min You’ll guess: On ShapeNet, how many ZipTok3D tokens did it take to reach reconstruction quality comparable to the 32-token COD-VAE baseline?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.