pretty_name: MotionFix MotionHub Format
language:
- en
license: other
tags:
- motion
- motion-editing
- smpl-h
- motionhub
task_categories:
- text-to-3d
MotionFix MotionHub Format
This dataset contains the processed MotionFix motion-editing data in the MotionHub format. It stores paired source and target motions as SMPL-H 52-joint parameter files, plus editing instructions and split annotations.
Please also follow the license and terms of the original MotionFix dataset.
Structure
MotionFix/
├── smplh_52/
│ ├── train/
│ │ ├── 000000_000499/
│ │ ├── 000500_000999/
│ │ └── ...
│ ├── val/
│ └── test/
├── pairs/
│ ├── train/
│ │ ├── 000000_000499/
│ │ ├── 000500_000999/
│ │ └── ...
│ ├── val/
│ └── test/
├── motionfix_train.json
├── motionfix_val.json
├── motionfix_train_val.json
├── motionfix_test.json
├── motionfix_all.json
├── stats.json
└── stats_target_only.json
Splits
| Split | Pairs | Frames | Hours |
|---|---|---|---|
| train | 6,730 | 744,502 | 6.894 |
| val | 330 | 36,420 | 0.337 |
| test | 1,013 | 111,454 | 1.032 |
| train_val | 7,060 | 780,922 | 7.231 |
| all | 8,073 | 892,376 | 8.263 |
Each pair has one source motion and one target motion. The smplh_52/ directory therefore contains 16,146 .npz files.
Files are sharded by pair id ranges so that no directory exceeds Hugging Face's file-count limit.
For example, pair 000000 stores its motions under smplh_52/train/000000_000499/ and metadata under pairs/train/000000_000499/.
Annotation Fields
Each item in motionfix_*.json includes:
motion_path,smplh_path: target motion path.source_motion_path,source_smplh_path: source motion path.caption_path,edit_pair_path: pair metadata path.caption,instruction: editing instruction text.fps,duration,num_frames,language.
Each file in pairs/ includes the source/target motion paths, the edit instruction, source/target timestamps, and canonicalization metadata.
Motion Format
Each .npz stores SMPL-H compatible fields:
poses:(T, 156)axis-angle pose parameters.global_orient:(T, 3).body_pose:(T, 63).left_hand_pose,right_hand_pose:(T, 45)each.trans,transl:(T, 3)body-model translation.betas: zero-beta shape vector.expression: zero expression vector.mocap_framerate,gender,num_frames.
The processed motions are canonicalized to the MotionHub SMPL-H convention and use body-model translation.