Dataset Preparation

June 23, 2026 · View on GitHub

GEM uses preprocessed datasets from GVHMR. Download the *_hmr4d_support.tar.gz archives from GVHMR's Google Drive and extract them under inputs/:

inputs/
├── AMASS/hmr4d_support/
├── BEDLAM/hmr4d_support/
├── H36M/hmr4d_support/
├── 3DPW/hmr4d_support/
├── EMDB/hmr4d_support/
└── RICH/hmr4d_support/

By downloading these files you agree to the original dataset licenses. The preprocessed data is for research use only.

Evaluation Datasets

Required for running task=test:

DatasetArchiveConfig
EMDBEMDB_hmr4d_support.tar.gzconfigs/test_datasets/emdb1_v1_fliptest.yaml
3DPW3DPW_hmr4d_support.tar.gzconfigs/test_datasets/3dpw_fliptest.yaml
RICHRICH_hmr4d_support.tar.gzconfigs/test_datasets/rich_all.yaml

Supplemental EMDB / RICH / 3DPW Test Files

Some GEM-SMPL test dataloaders also require supplemental files that may be missing from the GVHMR archives:

inputs/EMDB/hmr4d_support/emdb_vimo.pt
inputs/EMDB/hmr4d_support/emdb_slam_traj.pt
inputs/RICH/hmr4d_support/rich_test_vimo_preproc.pt
inputs/3DPW/hmr4d_support/test_3dpw_vimo_labels.pt
inputs/3DPW/hmr4d_support/3dpw_test_slam_traj.pt

Download them from the GEM-X HuggingFace repository and merge the included inputs/ directory into the repository root:

hf download nvidia/GEM-X \
  --include "gem_smpl/missing_hmr4d_support/**" \
  --local-dir .

mkdir -p inputs
cp -a gem_smpl/missing_hmr4d_support/inputs/. inputs/

Training Datasets

Required for training gem_smpl_regression:

DatasetArchiveConfig
AMASSAMASS_hmr4d_support.tar.gzconfigs/train_datasets/amass_v11.yaml
BEDLAMBEDLAM_hmr4d_support.tar.gzconfigs/train_datasets/bedlam_v2.yaml
Human3.6MH36M_hmr4d_support.tar.gzconfigs/train_datasets/h36m_v1.yaml
3DPW3DPW_hmr4d_support.tar.gz (train split)configs/train_datasets/3dpw_v1.yaml

Additional Datasets

Required for the full model gem_smpl (regression + text/audio generation):

DatasetSource
AIST++https://google.github.io/aistplusplus_dataset/ — dance motion
Beat2https://pantomatrix.github.io/BEAT/ — gesture dataset
HumanML3Dhttps://github.com/EricGuo5513/HumanML3D — text-to-motion