When a language model writes a sentence, it works with tokens: small pieces that can be predicted, rearranged, and checked. A motion tokenizer does something similar for movement. It turns a continuous body sequence—joint positions, rotations, or contact events—into a compact sequence of discrete symbols, then reconstructs movement from those symbols.

Why compress movement?

Raw motion is expensive. A short recording may contain hundreds of frames, many joints, and several coordinates per joint. A tokenizer learns a smaller vocabulary that keeps the patterns useful for generation: a step, a reach, a turn, or a pause. Compression makes long sequences easier to model and gives a system a common interface for text, motion, and video.

The compression is not neutral. If the vocabulary is too coarse, it erases timing or contact. If it is too large, generation becomes slow and brittle. Good tokenizers therefore need reconstruction tests as well as language-style likelihood scores.

Why add video?

MoVT, a recent video-augmented motion tokenizer, uses visual evidence to help text-to-motion generation. Video contains cues that a motion-capture skeleton may omit: object interaction, scene rhythm, and the appearance of a transition. The aim is not to reproduce pixels, but to learn a representation that makes the resulting movement more coherent.

That idea connects to ReMoMask-2, which retrieves structurally related motion before filling masked parts, and Open-UniMo, which gives motion a large token vocabulary inside a language model. In all three cases, the representation is doing editorial work: it decides which differences are meaningful before the generator starts improvising.

What should a reader look for?

Ask three simple questions. Can the representation be decoded without visible artifacts? Can it preserve contact and timing, not only pose? And can a person steer it with words that match what the body actually does? A low error on average joints can still hide a broken handoff, an impossible foot slide, or a missed pause.

Tokenizers are therefore less like a magic compression button and more like a dictionary. The dictionary shapes what the system can say. Our latest research digest covers the physics and retrieval layers around that dictionary; the community notes on contact and shape show what the dictionary risks leaving out.

References