Neural Motion Blending Across Arbitrary Character Topologies

Luca Cazzola1 · Giulia Martinelli1,2 · Nicola Conci1,2
1 University of Trento · 2 CNIT · MMLab

🚀 Quick start

⚠️ Just want to try the tool out ??? Check out BlendAnything, the Blender plugin that brings this work straight into Blender's NLA editor. ⚠️

1 · 🛠️ Setup

Get the code

git clone https://github.com/mmlab-cv/neural_motion_blending.git
cd neural_motion_blending

Dependancies and environment

conda env create --file environment.yaml
conda activate neural_motion_blending
pip install --no-build-isolation git+https://github.com/inbar-2344/Motion.git

(The classical NLA baseline and Blender visualizer additionally need a Blender install with Motion importable from Blender's own Python.)

2 · 📦 Get the data & models

Truebones Zoo is a closed-source, paid asset — see docs/DATA.md to purchase and process it.

Download the pretrained blending checkpoints, blendany_model_weights.zip, and extract its model folders into save/. The model reported in the paper is truebones_globpool (save/truebones_globpool/{args.json,model*.pt}) — the "attention-pool" variant (truebones_attnpool) is also included as an alternative (see docs/TRAIN.md).

3 · 📖 Usage

Guide	Covers
docs/DATA.md	🦴 Preparing Truebones and custom skeletons
docs/TRAIN.md	🏋️ Training AnyTop / MoDiffAE
docs/SAMPLE.md	🎬 Generating, blending & retargeting motion (sample/mix.py), the NLA baseline
eval/README.md	📊 Blending & latent-FID benchmarks
docs/VISUALIZATION.md	🎥 Stick-figure renders

💜 Cite Us 💜

If you use this work, please cite the CGI 2026 paper. In the mean time... the ArXiv prepring.

@misc{cazzola2026neuralmotionblendingarbitrary,
      title={Neural Motion Blending Across Arbitrary Character Topologies}, 
      author={Luca Cazzola and Giulia Martinelli and Nicola Conci},
      year={2026},
      eprint={2607.10370},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.10370}, 
}