PerMo

Standalone PerMo release normalized to the repository SMPL-H convention.

Summary

PerMo is released here as a standalone motion dataset with SMPL-H motions, single-motion captions, and style editing / style transfer instruction pairs.

• Motions: 6,610 clips, 924,726 frames, 8.56 hours at 30 FPS.
• Train split: 6,543 clips, 915,162 frames, 8.47 hours.
• Test split: 67 clips, 9,564 frames, 0.09 hours.
• Editing train split: 6,176 Neutral-to-style pairs, 869,194 target frames, 8.05 target-motion hours.
• Style groups: Age, Character, Condition, Emotion, Surroundings, Traits.

Layout

• smplh_52/: SMPL-H 52-joint axis-angle motions, canonicalized with the SMPL-H floor convention.
• hierarchical_caption/: per-motion hierarchical captions.
• augmented_caption/: PerMo action/style captions used by HYMotion M2M caption training.
• editing_caption/: Neutral-to-style instruction captions used by HYMotion M2M editing/style transfer training.
• pairs/: source/target pair metadata for PerMo style transfer.
• qwen3_augmented/, qwen3_editing/: pre-extracted text embeddings matched to the caption directories above.
• train.json, test.json, all.json: metafiles with paths rooted at PerMo/.
• editing_train.json: PerMo style editing/transfer pairs with source_motion_path and SMPL-H targets.
• stats.json: normalization statistics computed over train.json + test.json.

The old nested PerMo/20260513/ wrapper and 198-dim cache arrays are intentionally not part of this release.

Caption Sets

hierarchical_caption/ follows the MotionHub standard schema with macro, meso, and micro caption fields. augmented_caption/ preserves the original HYMotion-compatible short-caption format and is paired with qwen3_augmented/. editing_caption/ stores instruction-style captions for real source-to-target style editing pairs and is paired with qwen3_editing/.

Path Convention

Metafiles use paths rooted at PerMo/. For HYMotion training configurations whose data_dir is data/motionhub, the corresponding combined annotation uses ../PerMo/... paths.