Distribution quality
Official normalized kinetic and geometric feature spaces
Music-to-Dance · AIST++
Music-conditioned dance generation on the public AIST++ cross-modal protocol. Official kinetic and geometric metrics are reported beside normalized uTMR FID, time-domain beat alignment, and shared SMPL-22 physical diagnostics.
Task: Music-to-Dance
Dataset: AIST++
40 cross-modal cases
Audio-synchronized SMPL Mesh
GT excluded from ranking
Current Public SnapshotLoading results
Ranked methods
1
Released checkpoints only
Evaluation cases
40
All cases available in 3D
GT reference pool
1,320
Valid AIST++ motions
Motius joint protocol
30 fps
Canonical SMPL-22 joints
Method Comparison
Reference and generated distributions under the official and Motius evaluators.
Motius motion quality
L2-normalized universal TMR motion embedding FID
Music synchronization
Time-domain BeatAlign on standardized 30 fps motion
All-Case Motion Comparison
GT, Bailando, EDGE, TM2D, and UniMuMo native 24-joint skeletons overlaid on their SMPL Meshes, synchronized to AIST audio.
Results
Lower is better for FID and physical error; higher is better for diversity and BeatAlign.
| Method | Official Bailando protocol | Motius protocol | Physical diagnostics | ||||||
|---|---|---|---|---|---|---|---|---|---|
| FID_k ↓ | FID_g ↓ | Div_k ↑ | Div_g ↑ | uTMR FID ↓ | BeatAlign ↑ | Jitter ↓ | Float ↓ | Slide ↓ | |
Official feature metrics use the public Bailando 60 fps protocol. uTMR FID and ranked BeatAlign use canonical SMPL-22 joints at 30 fps while preserving the original timing in seconds.
Evaluation Contract
Paper comparability and cross-method fairness are reported as distinct protocol views.
Official quality
First-root-anchored SMPL-24 joints. FID_k and FID_g use 1,320 public AIST++ reference motions and GT-derived normalization.
Motius FID
Canonical SMPL-22 joints66 at 30 fps, capped at 20 seconds, encoded by the universal joint-position TMR and L2-normalized before FID.
Beat alignment
Motion beats are speed minima after 83.3 ms Gaussian smoothing. Music-to-motion distance uses a 50 ms kernel in seconds.
Physical motion
Jitter, floating, and sliding are measured on the shared 22-joint body and reported as diagnostics.
Paper Parity
Paper-reported values are audit targets and never participate in ranking.
| Row | FID_k | FID_g | Div_k | Div_g | BeatAlign |
|---|
Loading BeatAlign audit.