Generating Diverse and Natural 3D Human Motions from Text

How to read this

Category

Dataset and method: text-to-3D-human-motion generation

Contributions

• Introduces HumanML3D, a large-scale dataset of 14,616 motion clips with 44,970 text descriptions
• A two-stage pipeline: text2length sampling of motion length conditioned on text, then a temporal VAE that synthesizes motions
• Operates on motion-snippet codes capturing local semantic contexts to produce diverse, variable-length motions more faithful to text than deterministic baselines

Context

Sits in the text-conditioned human-motion generation area, contributing a new large dataset alongside KIT-ML and a stochastic VAE alternative to deterministic text-to-motion methods.

Correctness

Claims of greater diversity and faithfulness are evaluated on HumanML3D and KIT-ML against deterministic baselines, so results are tied to those datasets and metrics; the snippet-code representation and learned length distribution are assumptions whose effect on out-of-distribution prompts is not the focus.

Clarity

Accessible; a first pass conveys the dataset and two-stage idea, with a second pass for the VAE and snippet-code details.

How to read it

First pass for the dataset scale and the text2length-then-text2motion structure; second pass if you plan to use HumanML3D or build on the temporal VAE, noting the dataset is a contribution in itself.