FloodDiffusion 2
Efficient and Path Controllable Streaming Motion Generation
FloodDiffusion 2 supports continuous text-conditioned motion generation with Partial Attention, diffusion-compatible geometric supervision, and root-path control. This release includes six FD2 checkpoints and the paired 263D LDF/VAE models for first-generation compatibility.
| Checkpoint folder | Configuration | Training step | CFG |
|---|---|---|---|
humanml3d_fk_60k |
df_humanml3d_263.yaml |
60,000 | 4 |
humanml3d_path_fk_55k |
df_humanml3d_263_path.yaml |
55,000 | 3 |
humanml3d_babel_fk_200k |
df_humanml3d_babel_263.yaml |
200,000 | 4 |
humanml3d_babel_path_200k |
df_humanml3d_babel_263_path.yaml |
200,000 | 4 |
seed_fk_300k |
df_seed_138.yaml |
300,000 | 2 |
seed_path_fk_300k |
df_seed_138_path.yaml |
300,000 | 2 |
ldf_263 |
ldf_263.yaml |
205,000 | 5 |
vae_263 |
vae_263.yaml |
2,250,000 | — |
Each checkpoints/<folder>/ contains model.ckpt, a matching config.yaml,
and normalization statistics under assets/. FD2 folders also include the
quadratic FK matrix assets/W.npy. Checkpoints retain model, EMA, optimizer and
scheduler state.
Setup
Clone the code repository and run the following in a Python 3.10+ environment:
python setup_project.py
The command installs Python requirements and downloads all eight checkpoints,
their assets, and the original
FloodDiffusion dependency archive
for UMT5, T2M and GloVe. Add --with-data to also install the
prepared datasets
and paired VAE tokens.
For mesh rendering, obtain SMPL-H from its official site and import the licensed neutral model:
python setup_project.py --smplh /path/to/neutral/model.npz
Evaluation
python evaluate.py --config configs/df_humanml3d_263.yaml
The selected YAML defines the checkpoint, dataset and evaluation settings. Use the corresponding configuration for other variants. LDF uses the paired VAE from this release.
Dependencies
Third-party weights and datasets retain their respective licenses. SMPL-H is obtained separately from its official distributor.