MoSAIC: Aligned Intervention Supervision for Part-Local Motion Style Transfer

Two-reference composition. The content motion supplies the shared action and trajectory, hands-high supplies the arm style, and open-leg supplies the leg style. MoSAIC combines both references in one coordinated output at guidance scale 3.25.

Abstract

Editing character motion often requires borrowing an arm gesture or leg gait from reference sequences while retaining the original action, timing, root path, and unselected body regions. Motion datasets rarely provide paired targets for arbitrary local content–reference combinations, and self-reconstruction can let a diffusion model copy content without learning to respond to local references. We present MoSAIC for controllable character animation. The latent diffusion framework factorizes content and style by body part, conditions root trajectory separately, and uses a user-specified map to route references to anatomical regions. To make this behavior learnable, aligned intervention supervision applies a known local transformation to a source motion and constructs synchronized references and counterfactual targets that distinguish the requested response from what should remain unchanged. Anatomically constrained injection then supplies the routed cues during denoising. The validation-selected model retains HumanML3D generation quality. In a frozen evaluation of 128 motions and 896 routed conditions, masked routing achieves a fractional selected-target improvement of 0.198. Compared with whole-body routing, it reduces preserved-region error from 70.64 to 66.45 mm and matched-noise off-target leakage from 18.08 to 9.88 mm. MoSAIC therefore provides a more favorable response–preservation trade-off for part-local editing. Arm responses remain weaker than leg responses, and quantitative population-level natural-reference and multi-reference generalization remain unevaluated.

Method

Part-factorized conditioning

Content and style conditions retain their anatomical indices, with the root trajectory represented separately.

User-specified routing

A source map selects content or a reference motion independently for each anatomical region.

Aligned intervention supervision

Known local transformations create synchronized references and counterfactual targets, identifying both the requested response and the motion that should be preserved.

Masked response–preservation control

Reference cues are routed to selected regions while unselected regions receive content-derived conditions, limiting off-target influence.

Three rows of overlaid motion poses comparing factual motion, an aligned transformed reference, and a counterfactual target for arms, legs, and spine interventions — **Aligned intervention supervision.** A controlled transformation produces a synchronized reference, while applying the same transformation only to the routed region produces the counterfactual target. Factual motion, reference, and target share action, timing, sequence length, and root trajectory.

Main Results

On the frozen evaluation of 128 motions and 896 routed conditions, masked routing improves the balance between local response and preservation:

0.198 masked selected-target improvement (STI).
Preserved-region error (PRE) decreases from 70.64 to 66.45 mm relative to whole-body routing.
Off-target leakage (COL) decreases from 18.08 to 9.88 mm relative to whole-body routing.
42.4% of total influence lies on the requested-route diagonal.

Arms remain weaker than legs and spine. These results show a better response–preservation balance; they do not establish overall state-of-the-art motion style transfer.

Motion sequences comparing content, style reference, whole-body transfer, and selected-part transfer for arms, legs, and spine routing — **Part-local routing.** Rows route hands-high style to the arms, open-leg style to the legs, and old-motion style to the spine. Columns compare the content, reference, whole-body transfer, and selected-part transfer.

Video Results

Whole-body transfer. The reference is routed to all anatomical regions.

Arms-only transfer. Hands-high style is routed to both arms.

Legs-only transfer. Open-leg style is routed to both legs.

Two-reference composition. Hands-high supplies the arms; open-leg supplies the legs. Guidance scale: 3.25.

Citation

Please cite the arXiv preprint below.

@misc{amini2026mosaic,
  title         = {MoSAIC: Aligned Intervention Supervision
                   for Part-Local Motion Style Transfer},
  author        = {Amini, Nazanin and Desai, Kevin},
  year          = {2026},
  eprint        = {2607.26304},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2607.26304}
}

Acknowledgments

This work was partially supported by the U.S. National Science Foundation under Award Nos. 2211785 and 2316240.