Human motion generation goes image diffusion: joint trajectories become images, and two skeletons share one model

Text-conditioned 3D human motion can be generated through image diffusion: root-centered joint trajectories are ordered by body part and kinematic chain into a joint-time image, turning motion into something image generative architectures can process directly. Such models previously relied on vector representations tied to a particular skeleton, and HumanML3D and KIT-ML differ in joint counts and indexing conventions, so vector dimensionality and the joint each entry corresponds to both change, requiring explicit handling to share inputs and outputs across datasets. On HumanML3D the dataset-specific model raises Top-1 R-Precision from 0.500 for MARDM-SiT-XL to 0.540 and lowers FID from 0.114 to 0.089; on KIT-ML FID drops from 0.242 to 0.171 and Top-1 R-Precision rises from 0.387 to 0.395.

Global root motion stays out of that image and travels through a separate root-motion latent stream, and local-motion decoding is conditioned on the root latent so whole-body displacement and local articulation inform each other during generation. A common image shape alone does not encode anatomical relationships or capture dependencies between global movement and local articulation, so root motion is modeled separately and coupled to image decoding. Ablations show that removing root-conditioned decoding raises FID from 0.106 to 0.116 and lowers Top-1 R-Precision from 0.533 to 0.525, while in shared training the model without a dataset condition reaches FID 0.278 versus 0.093 with it.

One motion autoencoder and one text-conditioned generator can serve both skeletons without skeleton retargeting and without padding joints to a common count. Earlier generation across heterogeneous skeletons either retargeted motion or synthesized under skeleton conditioning, and sharing parameters usually required a unified joint layout. Shared training improves all six reported metrics on both benchmarks, lowering FID from 0.089 to 0.082 and raising Top-1 R-Precision from 0.540 to 0.546 on HumanML3D, and lowering FID from 0.171 to 0.145 and raising Top-1 R-Precision from 0.395 to 0.405 on KIT-ML.

The report covers text-to-motion generation on the HumanML3D and KIT-ML benchmarks, with motion processed at 20 FPS and 12.5 FPS under each dataset's native skeletal definition and standard training, validation and test splits. It is aimed at researchers and engineers who need to share a motion generation model across heterogeneous skeletons; evaluation follows the MARDM protocol and its pretrained evaluators, repeated 20 times with mean results and 95% confidence intervals. The authors propose extending the formulation to a broader range of skeletal structures and motion datasets.

The shared model's gains depend on the dataset condition, and without it shared training reaches FID 0.278, worse than the dataset-specific baseline; the root flow-matching weight changes performance only slightly over the tested range. The authors list validation on a broader range of skeletal structures and motion datasets as future work, so the current conclusions are limited to the two reported benchmarks.