Deep Motion Editing

Use this repo skill when a task names Deep Motion Editing, BVH motion editing, skeleton-aware retargeting, unpaired motion style transfer, OpenPose motion style, Mixamo BVHs, or Blender rendering/skinning. It is a script-oriented research checkout, not a conventional installable Python package. Read references/repo-provenance.md before assuming the skill matches a new checkout.

Start safely

This source snapshot has no setup.py, pyproject.toml, requirements file, or console entry point; do not invent pip install deep-motion-editing. For the selected Python workflows, install only the route dependencies in a disposable environment, for example:

python -m pip install "numpy<2" scipy pyyaml tqdm torch
# Add these only for training/logging/probe work:
python -m pip install tensorboardX tensorboard matplotlib scikit-learn

Use Python 3.7+ as documented by the project. CUDA is recommended for model inference/training; CPU is useful for format checks but is only a partial neural substitute. Use Blender 2.80-era or a tested newer Blender for bpy workflows; ordinary Python cannot import bpy. Keep datasets, checkpoints, OpenPose outputs, FBX assets, and generated results outside the skill tree. Do not download or overwrite them implicitly. Run the read-only environment probe scripts/check_environment.py before a new task. Read references/troubleshooting.md for legacy imports, paths, optional dependencies, and backend failures.

Route to the focused workflow

User intent	Read next
BVH parsing/writing, topology, JSON validation, kinematics, cleanup	sub-skills/animation-data/SKILL.md
Mixamo/custom data, intra/cross retargeting, evaluation, training	sub-skills/motion-retargeting/SKILL.md
3D/2D style transfer, normalization, Xia/BFA preparation, training	sub-skills/motion-style-transfer/SKILL.md
Blender BVH load/render, FBX skinning or conversion	sub-skills/blender-visualization/SKILL.md

The central cross-workflow contract is: validate BVH hierarchy and frame data first, preserve source inputs, write new outputs, then validate the result before neural comparison or rendering. Sibling routes link to each other when a task crosses that boundary.

Verification boundary

A parser/help check does not reproduce a paper result. Meaningful neural runs need the correct pretrained checkpoints, normalization files, dataset layout, model configuration, and a suitable backend. Blender routes additionally need a Blender executable and compatible FBX assets. This generated skill was built from a 2021 source snapshot and intentionally omits large binaries and external data; those limitations are recorded in the references and external verification artifacts, not hidden.