Timo

Timo: Taming Multimodal Diffusion Transformer 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 fuse text and motion through a pooled caption or cross-attention, rather than modeling both at the token level. Multimodal diffusion transformers do that joint modeling for images and video. Applied directly to motion, they still leave a gap: framewise reconstruction does not by itself produce coherent changes in body orientation.

We propose Timo, a kinematics-aware multimodal diffusion transformer for human motion. It combines fully shared text–motion attention with flow matching, geometric and rotational-kinematics supervision, and a curriculum that goes from broad motion learning to detailed caption alignment.

On the six-axis comparison below, Timo's average is 86.6, 25.1 points above the next system (61.5). Text-Motion Match gains 34.2 points over the strongest baseline on that axis, and Smoothness 10.9. Kimodo remains highest on Plausibility.

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 the clean sample, not a velocity field. **(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.

Shared text and motion

Text and motion share one transformer in every block: the same attention, MLP and adaLN weights. Flow matching predicts the clean pose, so eight solver steps suffice.

Kinematics-aware losses

Reconstruction is blind to frame-to-frame jitter. We also supervise decoded geometry and rotational kinematics, then train from broad motion to detailed captions. The architecture stays unchanged.

One six-axis scale

Four released systems are rescored on the same 0–100 axes. Timo's average is 86.6, 25.1 points above the next result. Kimodo remains highest on Plausibility.

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 averages 86.6, 25.1 points above 61.5; Text-Motion Match is 34.2 points above the best other score on that axis, and Smoothness 10.9. Kimodo remains highest on Plausibility.

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 Transformer for Human Motion Generation},
  author  = {Wang, Zhao and Hu, Jiangtao and Yu, Jack and Yu, Tao},
  journal = {arXiv preprint arXiv:XXXX.XXXXX},
  year    = {2026}
}