Timo

Timo: Taming Multimodal Diffusion Transformers for Human Motion Generation

Zhao WangJiangtao HuJack YuTao Yu^†*

LimX Dynamics ^†Project Lead ^*Corresponding Author

One request, twelve prompts, 36 seconds, generated as a single motion. Training windows are 8 seconds long; sampling tiles them and fuses the overlaps inside every denoising step.

0hours of motion

0.49Bparameters

0sampling steps

0evaluation axes

0physical humanoids

01Abstract

Most human motion generators inject text through cross-attention, where motion reads fixed text features but does not update them in return. Multimodal diffusion transformers instead let text and motion tokens refine each other, making their transfer to motion a natural next step.

Articulated motion is temporally coherent but weakly correlated across joints. A direct MMDiT transfer with flow matching produces poorly coordinated, jerky motion: neighboring joints need not share a pose, but their changes over time must remain coordinated.

We therefore propose Timo, a kinematics-aware multimodal diffusion transformer for human motion. Fully shared attention updates text and motion bidirectionally; geometric and rotational-kinematics losses supervise the articulated body and its evolution; and a two-stage curriculum moves from broad motion learning to detailed caption alignment. Timo surpasses Kimodo on five of six axes, with a 40.8% relative improvement in the average score.

02Method

Timo architecture and training objective — **Kinematics-aware multimodal generation.** **(a)** Text and motion tokens enter one 24-block stack as a single sequence predicting clean body parameters directly. **(b)** A block shares QKV, MLP and adaLN weights across the two modalities. **(c)** Reconstruction, rotational kinematics and decoded geometry. At right, two predictions with equal reconstruction loss but different temporal errors; only L_kin sees the difference.

A kinematics-aware MMDiT

Fully shared attention updates text and motion bidirectionally, while flow matching generates explicit body parameters directly.

Coherent, text-aligned motion

Geometric and rotational-kinematics supervision constrains poses and their evolution. A two-stage curriculum moves from broad motion learning to detailed caption alignment.

A six-dataset benchmark

One common six-axis evaluation compares Timo with four released systems and examines the framework through correlation analysis and training ablations.

03Six-axis evaluation

Each axis is a custom 0–100 score, higher better, so the columns share one scale. Text-Motion Match, Naturalness and Diversity are scored in a contrastive text–motion embedding. Smoothness uses third-order differences of the body-frame angular increment. Plausibility is limb-capsule interpenetration, a geometric proxy rather than a test of balance or dynamics. Speed combines time to the first output and per-segment time at batch size one. The last column is the mean of all six. Timo surpasses Kimodo on five of six axes and raises the average from 61.5 to 86.6, a 40.8% relative improvement. Kimodo remains highest on Plausibility (93.4 vs. 85.0).

Method	Institute	Params	Nat.	Match	Div.	Smth.	Speed	Plaus.	Avg
MotionMillion	Shanghai AI Lab (2025)	7.8B	42.3	30.5	51.7	49.5	45.5	72.1	48.6
HY-Motion	Tencent Hunyuan (2025)	1.0B	38.0	52.6	67.3	54.3	31.4	71.4	52.5
GENMO	NVIDIA (2025)	523M	44.2	28.0	52.3	79.8	20.6	68.1	48.8
Kimodo	NVIDIA (2026)	283M	36.8	51.0	59.1	79.0	49.5	93.4	61.5
Timo	LimX Dynamics (2026)	495M	86.4	86.8	89.3	90.7	81.4	85.0	86.6

04Against prior work

The same prompt through five models, in one camera and one crop. The pickup clips differ in how far the body descends and whether it stands back up; the gesture clips differ in whether the arm stays extended.

05Motion of any length

A longer request reuses the same generator.

Training sees 8-second windows. A longer request is tiled with overlapping windows, and those overlaps are blended inside every sampling step, without growing the attention context.

The clips below compose several prompts into one motion, up to 36 seconds, mixing locomotion with gestures and posture changes. They illustrate that composition. They do not measure how far the motion drifts.

The overlay reports the active prompt and the tiling boundaries.

06Executed on hardware

The generator supplies a kinematic reference. Each sample is retargeted and tracked on two machines, LimX Luna and LimX Oli. These clips show that the same outputs can be executed on both. They are not a success rate across robots or prompts.

07Further generations

One checkpoint, one sampling budget, no per-prompt tuning. Drag or scroll sideways; click any clip to enlarge.

08Training data

Training data distribution — **Corpus composition.** **(a)** Motion categories over all 1,983 hours; the three rings are family, subfamily and action, and arc is hours. **(b)** Caption length, shown for whole takes, one-clip recordings, and VLM-labelled clips. **(c)** Caption vocabulary.

09BibTeX

@article{timo2026,
  title   = {Timo: Taming Multimodal Diffusion Transformers for Human Motion Generation},
  author  = {Wang, Zhao and Hu, Jiangtao and Yu, Jack and Yu, Tao},
  journal = {arXiv preprint arXiv:2609.30761},
  year    = {2026}
}