Modelpublic

ZeyuLing/motius-evaluator-universal-smplh-joints66

sourceHugging Facemitupdated 2mo agoView on Hugging Face

0likes17downloads

Model Card

Motius Universal SMPL-22 Joints66 Evaluator

Motius reproduction of the TMR architecture, trained from scratch on the full HYMotion Data SFT set and the full single-person MotionHub training union. Motion is represented as canonicalized SMPL body-joint positions (22 joints in xyz) at 30 fps. This is a Motius-trained checkpoint, not an official TMR checkpoint.

Provenance

—Architecture paper: TMR: Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis
—Original architecture repository: Mathux/TMR
—Motius implementation: TMRBundle / TMRTextMotionEvaluator

Artifact layout

—model.safetensors: TMR motion/text projection encoders and reconstruction decoder
—text_encoder/: bundled DistilBERT tokenizer and weights for offline loading
—stats/: training-set joints66 normalization statistics
—config.json and preprocessor_config.json: architecture and protocol
—artifact_manifest.json: file and source-checkpoint hashes

The public checkpoint is epoch 248 from the corrected canonicalization, caption, training-data, and normalization run used by the Motius T2M leaderboard.

Download

python

from huggingface_hub import snapshot_download

artifact = snapshot_download("ZeyuLing/motius-evaluator-universal-smplh-joints66")

python

from motius.evaluation.evaluators import TMRTextMotionEvaluator

evaluator = TMRTextMotionEvaluator.from_pretrained(artifact, local_files_only=True)