title: Sequential Text-to-Motion · BABEL
emoji: 🎬
colorFrom: green
colorTo: indigo
sdk: static
app_file: index.html
pinned: false
license: mit
Sequential Text-to-Motion · BABEL
Static leaderboard for text-guided sequential human motion generation on the processed official BABEL validation split: 1,295 episodes, 7,285 LLM-rewritten action intervals, and 5,990 transition boundaries. Explicit transition labels are cut at their midpoint, and adjacent short actions are merged to at least 30 frames before caption rewriting.
Every method is converted to neutral zero-beta SMPL-22 joints. Semantic and transition distribution metrics use the Motius Joint-Position Evaluator. FID is measured only after per-sample L2 normalization of the uTMR motion embeddings; Peak Jerk and AUJ are computed directly from paired episode trajectories. GT/reference rows are shown for calibration and excluded from generated-method ranking. Results from the superseded 64-composition protocol are not carried into this leaderboard.
| Method | R@1 | R@2 | R@3 | Normalized FID | MM-Dist | Normalized Transition FID | AUJ Gap |
|---|---|---|---|---|---|---|---|
| BABEL GT | 0.3614 | 0.5284 | 0.6317 | 0.0000 | 47.8378 | 0.0000 | 0.0000 |
| MotionCanvas | 0.4240 | 0.6049 | 0.7134 | 0.0587 | 46.2662 | 0.0481 | 80.0543 |
| FlowMDM | 0.2504 | 0.3925 | 0.4818 | 0.0467 | 50.8503 | 0.0555 | 34.4040 |
| MotionStreamer | 0.2130 | 0.3303 | 0.4175 | 0.0610 | 52.0339 | 0.0702 | 76.2889 |
| MotionLab | 0.2580 | 0.3793 | 0.4536 | 0.2011 | 51.3873 | 0.2499 | 25.7259 |
| PRISM (epoch 26) | 0.2833 | 0.4314 | 0.5168 | 0.0591 | 51.0135 | 0.0739 | 157.8457 |
R-Precision uses official BABEL act_cat action-group multi-positive recall
batches of 32 (7,264 paired segments). The 7,285 intervals form 1,738 action
groups; synonymous labels are retained and never treated as false negatives.
Distribution metrics use all 7,285 segments. Encoder forward batch size is a
throughput setting independent of the 32-sample recall candidate batch. GT is
excluded from ranking.
PRISM uses checkpoint-epoch_26 with a fixed 360-frame canvas, CFG 5.0, and
AR5 for every model call; the complete 1,295-episode run contains no legacy
365-frame calls.
MotionLab uses its official five-frame autoregressive context. Both rows are
converted to the same canonical SMPL-22 joints before evaluation.
MotionCanvas uses checkpoint-epoch_1964, CFG 5.0, and five conditioned
frames between adjacent actions. Its final joints use a fixed zero-phase
three-frame [1, 2, 1] / 4 filter followed by canonical SMPL-22 bone-length
projection; the frozen 128-episode control changes semantic FID by only
+0.0002 while reducing Peak Jerk from 635.6 to 162.4.
The Space also includes a synchronized Three.js neutral-SMPL-mesh comparison
of BABEL GT, MotionCanvas, FlowMDM, MotionStreamer, PRISM, and MotionLab. FlowMDM and
MotionStreamer are fitted from the same native joints66 sequences used by
evaluation; PRISM and MotionLab use their fitted SMPL parameters. Each episode is
canonicalized once at frame zero to the same +Z-facing convention while its
subsequent global XZ trajectory and inter-subclip continuity are preserved.
The all-case explorer covers all 1,295 episodes, supports caption and episode-ID
search, and displays every retained method output together on one synchronized
timeline. Binary motion chunks
are loaded lazily from
ZeyuLing/Motius-Leaderboard-Cases.