Hocassian/Gitruck-MotionIR

Gitruck MotionIR

Gitruck MotionIR is a Chinese motion-design dataset that aligns project-level natural-language descriptions, technique-level annotations, temporal evidence, and a renderable intermediate representation (IR v1). The corpus was normalized from authorized Alight Motion, After Effects, NodeVideo, and Jianying projects.

Gitruck MotionIR 是一个中文动效设计数据集，将工程级描述、技法级标注、时间证据与可渲染 IR v1 对齐。语料由已获授权的 Alight Motion、After Effects、NodeVideo 与剪映工程归一化而来。

Dataset summary / 数据概览

Formats: AM 1,308, AE 487, NV 212, Jianying 88.

Configurations

• —technique_instances (default): one row per technique occurrence, with prompt, mechanics, parameters, evidence layers/effects, confidence, and time window.
• —projects: one row per project, with canvas, tags, prompt, statistics, and asset links.
• —ir_v1: one row per project containing the normalized renderable IR JSON and structural counts.
• —taxonomy: the 10-category, 220-pattern controlled vocabulary.

from datasets import load_dataset

techniques = load_dataset("Hocassian/Gitruck-MotionIR", "technique_instances")
projects = load_dataset("Hocassian/Gitruck-MotionIR", "projects")
ir = load_dataset("Hocassian/Gitruck-MotionIR", "ir_v1")
taxonomy = load_dataset("Hocassian/Gitruck-MotionIR", "taxonomy")

Splits

Projects are assigned deterministically by a salted SHA-256 hash of project_uid. All rows belonging to one project remain in the same split, preventing project-level leakage between training and evaluation data.

Renderable IR

IR v1 represents a beat-sized scene graph with canvas metadata, nested layers, per-channel keyframes, easing, and browser-native effect operators. The complete contract is in schema/IR_v1_spec.md. ir_json is stored as a string so that operator-specific effect parameters remain lossless and schema-stable in Parquet.

Preview assets

The repository includes deterministic archives rather than thousands of loose files:

• —assets/particles.tar.gz: 9.84 MiB
• —assets/posters.tar.gz: 17.20 MiB
• —assets/blueprint.tar.gz: 0.05 MiB

Extract all three archives into the same directory to preserve the relative references used by particle HTML files:

particles/
posters/
materials/blueprint/

Each project row exposes particle_archive_member and poster_archive_member. The particle HTML files use bundled GSAP and programmatically generated blueprint placeholders; no original image or video media is included.

Collection and annotation

Source projects were parsed deterministically into a shared analysis representation and then compiled into IR v1. Technique descriptions and mechanics were generated with model assistance, normalized against a controlled taxonomy, and validated programmatically. Public titles are stable pseudonymous identifiers. Text layers, free-form annotations, evidence strings, and particles pass the release sanitizer.

Quality and limitations

• —Confidence distribution: high 6,380, mid 3,116, low 371.
• —1,080 technique rows do not have a valid explicit time window; use has_time_window before consuming time_start_ms and time_end_ms.
• —IR v1 is a renderable structural approximation, not a pixel-identical copy of the source editor project. Some editor-specific easing and effect parameters are approximated.
• —Annotations are primarily Chinese and may contain model-generated errors. Use confidence for filtering and report corrections through the repository discussions.
• —The dataset contains motion structure and blueprint previews, not the original media assets.

Intended uses

Suitable uses include text-to-motion-IR generation, motion-technique classification, structured animation retrieval, editor-to-neutral-IR research, and controllable motion-graphics generation. Users remain responsible for validating generated output.

License and attribution

The dataset, annotations, normalized IR, and bundled preview assets are released under Creative Commons Attribution 4.0 International. Please attribute Gitruck MotionIR contributors and link to this repository.

Citation

@dataset{gitruck_motionir_2026,
  title  = {Gitruck MotionIR},
  author = {Gitruck MotionIR Contributors},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Hocassian/Gitruck-MotionIR}
}