| license: other | |
| tags: | |
| - motion-generation | |
| - flow-matching | |
| # Motion flow demo — checkpoints | |
| Weights and dataset statistics for the motion flow demo Space. Layout: | |
| | path | what | | |
| |---|---| | |
| | `weights/rvq_vae_best.pt` | frozen RVQ-VAE (encoder + decoder) — required | | |
| | `weights/latent_best.pt` | LFM, the latent flow model | | |
| | `weights/direct_best.pt` | CDFM, the direct flow model — drives in-betweening, composition, keyframes | | |
| | `weights/reflow_best.pt` | rectified student, usable at 1–4 steps | | |
| | `assets/Mean.npy`, `assets/Std.npy`, `assets/rest_len.npy` | HumanML3D statistics — required | | |
| | `assets/reference_clips.npz` | in-betweening reference clips | | |
| `z_mean` / `z_std` travel inside each latent checkpoint: they are the statistics of the | |
| RVQ-VAE latents as they stood when that model was trained, so a checkpoint loaded with the | |
| wrong pair decodes to noise. | |
| **No SMPL body model is included here.** SMPL is not redistributable; register at | |
| [smpl.is.tue.mpg.de](https://smpl.is.tue.mpg.de) to obtain it. | |