FineMotion-Style

FineMotion-Style accompanies Learning Kinematic Frequency-Aware Disentanglement for Motion Style Transfer and Editing, SIGGRAPH Asia 2026. The integrated research corpus uses six sources. This Hugging Face Dataset contains the two Bandai motion/text and BVH companion archives. Frozen protocol IDs are already included in the GitHub code repository. Other sources are acquired from their official providers and processed locally.

Project · Method code · Dataset code and reconstruction guide

Downloads

The selected components and their SHA256 values are recorded in the archive manifest.

• Bandai original v1 — motion, features and text: 3,077 archived SMPL24 source sequences, 2,676 paired 20Hz HumanML263 feature/text clips, source mappings, split projections and a complete BVH-to-MotionBuilder-to-SMPL example. The archived source sequences use 30Hz; another project can choose a different conversion rate.
• Bandai BVH companion — native and SMPL-derived views: 3,077 original animation BVH files and 3,077 matching SMPL-derived BVH exports, with explicit stage labels and source/clip mapping. These later SMPL-derived exports are not the original MotionBuilder output.

These two Bandai archives provide different representations of the same source collection; they do not reconstruct the full six-source research corpus. The previously prepared three-source text archive is not selected for this initial Hugging Face release.

Frozen protocol IDs on GitHub

The 292 original split/benchmark ID lists are available in the code repository: download the frozen protocol archive and read its package guide and checksum instructions. The file is 7,944,168 bytes, with SHA256 3209df832ac2729e30c1544bf1eb56e954179f4033efeb71a7588b27b4bd1a49. This GitHub package contains frozen memberships and package metadata, with no source captions, motions or model weights. Its IDs refer to all six sources; the protocol archive is supplied through GitHub separately from the two Hugging Face data downloads.

Acquiring the other sources

• HumanML3D, with AMASS where original motion parameters are needed.
• Motion-X.
• LAFAN1.
• Motorica Dance Dataset.
• FineDance.

Obtain each dataset, labels and required model assets under its provider's terms. The repository documents source conversion, ID/crop mappings, video descriptions, HumanML-format text, lemma/POS and benchmark rules. Historical code and recovered settings are distinguished from fresh end-to-end verification in the source guide.

Text, time tags and frozen IDs

HumanML3D and Motion-X use their inherited annotation routes. For the four newly annotated sources, historical MAIN descriptions commonly use source labels and DETAILS were generated with VideoLLaMA3. The text format is:

caption#lemma/POS lemma/POS ...#start_seconds#end_seconds

The two caption lines are alternatives for the same motion; 0#0 denotes the whole paired clip. POS tags describe words, not the four benchmark families.

Regenerating descriptions with a VLM need not reproduce the old wording or POS tokens. Keep a separately named text version while retaining the historical clip IDs, train/validation/test partitions and benchmark lists. Missing source files must be reported rather than filled with new IDs. Reclassifying revised captions defines a new benchmark version, not the frozen paper protocol.

The saved motion-availability pool has 59,038 IDs; the paper's 59,039 count has an unresolved one-ID discrepancy. The 67,907-ID base text/feature pool and its 13,357 strict style labels have different scopes. See the benchmark protocol.

Provenance, limitations and licences

Bandai data and annotations are distributed under CC BY-NC 4.0, with source attribution, original notices and the transformation history in the archive. The source creator is Bandai Namco Research Inc. The conversion uses our 2025 processing method. Project code, documentation and the protocol-ID material on GitHub retain their MIT notices; the card's data-licence tag does not replace component-specific terms. SMPL-Body rig attribution in the BVH view retains its separate CC BY 4.0 terms.

SMPL model weights, blend shapes and characterized target templates are obtained separately. Other providers' motions and full annotation corpora are not included. No source creator endorses this adaptation. See the source notices.

The original annotations are preserved; Qwen revision proposals are not applied automatically. The four-source first-pass review completed on 2026-10-16, but model flags are not ground-truth accuracy measurements. The selected October 16 Bandai packaging revision translates comments and removes absolute preparation paths while preserving all motion, feature, caption and split bytes from the October 14 archive. Current findings and bounded known issues are in the release guide.

Preserve an immutable original-v1 revision. Publish later accepted caption or conversion changes as separately identified versions with old/new hashes and change ledgers. A different text version does not silently replace frozen IDs.

Citation

@inproceedings{dong2026kinematic,
  title={Learning Kinematic Frequency-Aware Disentanglement for Motion Style Transfer and Editing},
  author={Dong, Ran and Xie, Haoran and Yang, Xi},
  booktitle={SIGGRAPH Asia 2026 Conference Papers},
  year={2026},
  note={to appear}
}

Retain the Bandai source credit and cite each additional source used in a local reconstruction; the FineMotion citation does not replace upstream attribution.