Text-to-Motion Leaderboard

HumanML3D official test split, selected GT caption protocol. Public rows report semantic retrieval metrics, universal SMPL-22 TMR metrics, and joint-level physical quality metrics. uTMR metrics use canonicalized SMPL-22 joints66 at 30 fps. Paper-only rows are kept separate when no released checkpoint can be evaluated through the shared SMPL-22 bridge.

Task: Text-to-Motion Dataset: HumanML3D official test MS samples: 4042 uTMR samples: 4034 Captions: selected full-clip GT

Current Public Snapshot Updated from verified internal metric JSONs. PRISM 1.0 is the no-KT/no-KAFS baseline rerun with pad360/crop; PRISM KAFS cfg5 uses the epoch-12 new-VAE checkpoint.

Best normalized uTMR FID

Generated methods only

Best MS R@1

Generated methods only

Best uTMR R@3

Generated methods only

Lowest Foot Slide

Generated methods only

Method Comparison

Best Second GT reference

Generated-method ranking

Normalized profile

100 is the best generated-method value on each axis; GT is excluded.

All-case motion comparison

4,042 selected-caption cases · lazy loaded

MS = MotionStreamer-272 evaluator. uTMR = universal SMPL-22 joints66 TMR evaluator. Only per-sample L2-normalized FID is displayed and ranked; historical raw-embedding FID values remain hidden until recomputed. Lower normalized FID and MM are better; higher R-Precision is better. Diversity is a reference statistic and is not ranked.

Slide, Float, and Jitter use the shared SMPL-22 joint-level physical metric implementation. PoseQ is the MBench NRDF pose-quality score. Dynamic is compared with the GT reference rather than minimized.

Paper benchmark rows reproduce metrics reported by the original papers when the released checkpoint is not available in the shared SMPL-22 evaluation protocol. They are not used for generated-method ranking.

Calibration rows quantify legacy SMPL-to-HML263 conversion loss and are not ranked as generation methods. uTMR sanity is tracked by the GT row in the semantic table.