1
stars
29
commits
1
linked in READMEs
Aug 31, 2026
updated
Unified public leaderboard for temporal-conditioned human motion generation on the HumanML3D official test split. The page exposes two protocols:
Every setting includes a visible GT reference row. GT semantic metrics are reused from the matching Text-to-Motion · HumanML3D benchmark rather than recomputed. For Temporal Control, condition error and failure rates are zero by identity, and foot skating is measured with one deterministic pass over the 4,012 temporal test clips. GT is excluded from ranks, best/second highlighting, and chart normalization.
The quantitative table includes all eight 4,012-case ProjFlow runs. ProjFlow's native Motius artifact and task APIs are documented in its Model Card. Its all-case mesh pack has not been added to the qualitative explorer, so the viewer does not present a partial or mixed-protocol ProjFlow row.
The page supports task filtering, method search, sortable metrics, bar and radar comparisons, and protocol-specific details. Ordered multi-prompt composition is maintained separately on the Sequential Text-to-Motion · BABEL benchmark.
Temporal Control also provides six all-case Three.js explorers. Each contains
all 4,012 official cases and the aligned GT, CondMDI, FlowMDM, KIMODO,
MotionLab, and MotionCanvas motions as synchronized neutral SMPL meshes rather
than joint-only skeletons. The selected setting can be switched
without loading unrelated motion chunks; large assets live in the public
ZeyuLing/Motius-Leaderboard-Cases
dataset.
The static Space is defined by:
index.html: page structure and responsive styling.leaderboard.js: interaction, ranking, and chart logic.temporal_control_results.json: structured Temporal Control result snapshot.cases/: one lightweight manifest and viewer per Temporal Control setting.29 commits
1
stars
29
commits
1
linked in READMEs
Aug 31, 2026
updated
Unified public leaderboard for temporal-conditioned human motion generation on the HumanML3D official test split. The page exposes two protocols:
Every setting includes a visible GT reference row. GT semantic metrics are reused from the matching Text-to-Motion · HumanML3D benchmark rather than recomputed. For Temporal Control, condition error and failure rates are zero by identity, and foot skating is measured with one deterministic pass over the 4,012 temporal test clips. GT is excluded from ranks, best/second highlighting, and chart normalization.
The quantitative table includes all eight 4,012-case ProjFlow runs. ProjFlow's native Motius artifact and task APIs are documented in its Model Card. Its all-case mesh pack has not been added to the qualitative explorer, so the viewer does not present a partial or mixed-protocol ProjFlow row.
The page supports task filtering, method search, sortable metrics, bar and radar comparisons, and protocol-specific details. Ordered multi-prompt composition is maintained separately on the Sequential Text-to-Motion · BABEL benchmark.
Temporal Control also provides six all-case Three.js explorers. Each contains
all 4,012 official cases and the aligned GT, CondMDI, FlowMDM, KIMODO,
MotionLab, and MotionCanvas motions as synchronized neutral SMPL meshes rather
than joint-only skeletons. The selected setting can be switched
without loading unrelated motion chunks; large assets live in the public
ZeyuLing/Motius-Leaderboard-Cases
dataset.
The static Space is defined by:
index.html: page structure and responsive styling.leaderboard.js: interaction, ranking, and chart logic.temporal_control_results.json: structured Temporal Control result snapshot.cases/: one lightweight manifest and viewer per Temporal Control setting.29 commits