RehabMotion FT Generator
Private research release of the official KIMORE v2 text-to-motion system. This repository contains the generator and the refined VAE that must be used together.
Checkpoints
| File | Description | Size | SHA-256 |
|---|---|---|---|
generator.ckpt |
MotionGPT3 generator selected at epoch 40 (update 4,920) | 2.49 GB | 46ad3b39607eaa4083ce107b9c5734fe2e8984f8256182bd8501b14467b814cf |
vae.ckpt |
Refined and latent-scale-calibrated MldVae | 72.2 MB | 446de7711e4cdff79ec8e08ceb15161f10e4293785462397d04397e633ddf46e |
Release identifier: kimore-v2-20260909.
What changed in v2
The VAE follows the MldVae Transformer architecture with 263-dimensional HumanML3D features, one 256-dimensional latent, nine layers, four attention heads, and a 1,024-dimensional feed-forward block.
The approved VAE lineage includes joint encoder/decoder adaptation followed by
10,000 additional encoder-only updates. Its latent scale was calibrated on the
training split using posterior means. The resulting inverse standard deviation
is 0.6275949456609311.
The generator was then trained with this VAE frozen. The selected generator is the epoch-40 checkpoint from a 100-epoch run, chosen by the lowest exercise-macro validation diffusion loss.
Expected inference settings
- Latent target: posterior mean
- Latent normalization: enabled
- Guidance scale: 3.0
- Temperature: 1.0
- Predict epsilon: enabled
- Diffusion batch multiplier: 4
- Motion features: 263
- Output skeleton: 22 joints
- Output frame rate: 20 FPS
- Rigidification: disabled
- Kinematic filtering: disabled
The generator and standalone VAE hashes should be verified before inference. Do not substitute another VAE or silently select a different generator checkpoint.
Download
from huggingface_hub import hf_hub_download
repo_id = "lucasbrandao/RehabMotion-FT-Generator"
generator_path = hf_hub_download(
repo_id=repo_id,
filename="generator.ckpt",
)
vae_path = hf_hub_download(
repo_id=repo_id,
filename="vae.ckpt",
)
print(generator_path)
print(vae_path)
The checkpoints use the project MotionGPT3/PyTorch Lightning model definitions; they are not standalone Transformers pipelines.
Associated dataset
Synthetic motions generated with this exact pair of checkpoints are stored in the private dataset repository RehabMotion-FT-Motions.
The original 1,000-motion release is under original_release/. Three paper generation banks, each with 1,000 motions, are under paper_banks/.
Intended use
This release is intended for research on text-to-motion generation and synthetic-data augmentation for rehabilitation exercise classification.
Limitations
- The model is not clinically validated and must not be used for diagnosis, treatment decisions, or patient monitoring.
- Visual approval and reconstruction metrics do not establish anatomical or clinical correctness.
- Generated motions can contain residual geometric and temporal artifacts.
- The training and evaluation data are specific to the five KIMORE exercise classes; behavior outside this domain is not established.
- Reproduction can vary across PyTorch and CUDA versions.