---
license: other
license_name: research-non-commercial
license_link: https://github.com/Coral79/FrankenMotion-Code/blob/main/LICENSE
tags:
  - motion-generation
  - human-motion
  - retrieval
  - smpl
  - cvpr2026
  - evaluation
library_name: pytorch
---

# FrankenMotion — Per-body-part Evaluation Encoders

Per-body-part TMR-style retrieval encoders used to evaluate motion generation in the CVPR 2026 paper *FrankenMotion: Part-level Human Motion Generation and Composition*.

- 📄 **Paper:** https://huggingface.co/papers/2601.10909
- 💻 **Code:** https://github.com/Coral79/FrankenMotion-Code
- 🤖 **Diffusion model:** [Coral79/frankenmotion](https://huggingface.co/Coral79/frankenmotion)
- 🧟 **Dataset:** [Coral79/frankenstein-dataset](https://huggingface.co/datasets/Coral79/frankenstein-dataset)

## What's inside

Nine independently trained TMR encoders (one per body part + caption + action), each pairing a motion encoder with a text encoder into a shared retrieval latent space:

```
action/, head/, left_arm/, left_leg/, right_arm/, right_leg/,
sequence_caption/, spine/, trajectory/
    ├── config.json
    └── last_weights/{motion,text}_encoder.pt
stats/{mean,std}.pt          # shared motion-feature normaliser stats
```

These encoders feed the `guo` and `guo+threshold` retrieval protocols (R@1, R@3, MM-Dist, FID, Diversity) reported in paper Table 1.

## Quick start

```python
from huggingface_hub import snapshot_download
snapshot_download(repo_id="Coral79/frankenmotion-eval-model", local_dir="pretrained/eval_model")
```

Then follow the [main repo README](https://github.com/Coral79/FrankenMotion-Code#-evaluation) — `src.eval.paper_table` auto-runs both retrieval protocols on CPU and reproduces the paper table.

## Citation

```bibtex
@inproceedings{li2026frankenmotion,
  title={{FrankenMotion}: Part-level Human Motion Generation and Composition},
  author={Li, Chuqiao and Xie, Xianghui and Cao, Yong and Geiger, Andreas and Pons-Moll, Gerard},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}
```