---
pretty_name: FloodDiffusion 2 Prepared Motion Data
language:
  - en
tags:
  - motion-generation
  - text-to-motion
  - humanml3d
  - babel
  - seed
---

# FloodDiffusion 2 Prepared Motion Data

Prepared representations, captions, and dataset splits used by the
[FloodDiffusion 2 code](https://github.com/AlayaLab/FloodDiffusion2).
[Model weights](https://huggingface.co/AlayaLab/FloodDiffusion2) are distributed separately.

| Archive | Representation | Motion sequences | Extraction directory |
|---|---|---:|---|
| `HumanML3D-263.zip` | HumanML3D, 263D | 29,228 | `data/HumanML3D/HumanML3D263/` |
| `BABEL-263.zip` | BABEL, 263D | 5,341 | `data/BABEL/HumanML3D263/` |
| `SEED-138.zip` | SEED, 138D | 142,220 | `data/SEED/MEI138/` |

All archives are created with `zip -0` (stored without compression). Extract them
from the code repository's root directory, or install everything with:

```bash
python setup_project.py --with-data
```

The setup command verifies SHA-256 checksums and extracts each archive to its
configured location. Use `python setup_project.py --check --with-data` to verify
the installed files. `data_manifest.json` pins the archive revision and checksums;
each dataset includes a `release_manifest.json` with per-file checksums.

```text
data/
├── HumanML3D/HumanML3D263/
│   ├── new_joint_vecs/
│   ├── texts/
│   ├── TOKENS_20260113_233726_vae_wan_z4_2250000/
│   ├── Mean.npy, Std.npy
│   └── split and token-statistics files
├── BABEL/HumanML3D263/
│   ├── new_joint_vecs/
│   ├── texts/
│   ├── TOKENS_20260113_233726_vae_wan_z4_2250000/
│   ├── Mean.npy, Std.npy
│   └── split and token-statistics files
└── SEED/MEI138/
    ├── new_joint_vecs_uni/
    ├── texts/
    ├── Mean_uni.npy, Std_uni.npy
    └── split files
```

The HumanML3D and BABEL archives include tokens from the released 2,250,000-step
VAE for LDF training. Caption lines preserve the `caption#tokens#start#end` format;
BABEL frame-level annotations are placed alongside the 263D motion features in
`texts/`. Original split memberships are retained. HumanML3D's two original
non-finite test sequences, `007975` and `M007975`, are retained in the archive and
skipped by the data loader.

Source datasets and their annotations retain their original terms. Cite the
corresponding source datasets as well as FloodDiffusion 2 when using these files.