All checkpoints are tracked in this repository: the benchmark checkpoints
reproduce the paper's numbers, and the deployment checkpoints run the streaming
and on-device demos right after cloning. All load with
flowavatar.models.load_checkpoint and training.evaluate. Benchmark files
carry their model_params, source experiment and epoch; deployment files
hold the state dict and epoch and take their architecture from the gru /
lstm preset.
checkpoints/
├── benchmark/ # SMPL+H, 16 betas, unified decoder: paper models
│ ├── gru_smplh.pt # paper Table 2 / 3 "FlowAvatar" (all losses)
│ ├── gru_smplh_no_temporal.pt # Table 3 ablation
│ ├── gru_smplh_no_hand.pt # Table 3 ablation
│ ├── gru_smplh_no_ground.pt # Table 3 ablation
│ └── lstm_smplh.pt # on-device architecture, benchmark setting: causal LSTM, 30 fps
├── GRU/baseline.pt # deployment: SMPL-X, 10 betas, streaming demo
└── LSTM/baseline.pt # deployment: SMPL-X, 10 betas, Unity Sentis export
Evaluate with
python -m training.evaluate --config training/configs/gru_smplh.yaml --checkpoint checkpoints/benchmark/gru_smplh.pt
python -m training.evaluate --config training/configs/lstm_smplh.yaml --checkpoint checkpoints/benchmark/lstm_smplh.ptThese need the SMPL+H body models and the prepared AMASS test data
(data/amass_smplh_60fps, and data/amass_smplh_30fps for the LSTM); see
training/README.md, sections 1–2. The metrics are also written
next to the checkpoint (<name>_eval_test.json; --output to change).
| Checkpoint | Config | MPJRE (deg) | MPJPE (cm) | MPJVE (cm/s) | Jitter (m/s³) | Hand PE (cm) | Paper |
|---|---|---|---|---|---|---|---|
gru_smplh.pt |
gru_smplh.yaml |
4.196 | 5.407 | 21.725 | 377.3 | 2.318 | Table 2: 4.19 / 5.40 / 21.72 / 3.77 / 2.31 |
gru_smplh_no_temporal.pt |
ablations/no_temporal.yaml |
4.555 | 6.083 | 25.588 | 517.4 | 2.506 | Table 3: 4.55 / 25.58 / 5.17 / 2.50 |
gru_smplh_no_hand.pt |
ablations/no_hand.yaml |
4.416 | 5.962 | 23.260 | 391.6 | 2.981 | Table 3: 4.41 / 23.26 / 3.91 / 2.98 |
gru_smplh_no_ground.pt |
ablations/no_ground.yaml |
4.313 | 5.558 | 22.551 | 405.3 | 2.256 | Table 3: 4.31 / 22.55 / 4.05 / 2.25 |
lstm_smplh.pt |
lstm_smplh.yaml (30 fps, T=20) |
4.299 | 5.905 | 25.419 | 199.4 | 2.340 | on-device (live demo) setting |
These are the outputs of training.evaluate on data prepared with
training.prepare_data (jitter in m/s³; the paper reports 10² m/s³ and truncates to
two decimals; its Table 3 lists MPJRE / MPJVE / Jitter / Hand PE). They
coincide with the research code's evaluation of the same files to the third
decimal. With the real-time protocol (--window 40, last frame of each
sliding window, as in the streaming demo) gru_smplh.pt gives MPJRE 4.239,
MPJPE 5.627, MPJVE 27.349 and jitter 788; lstm_smplh.pt with --window 20
gives 4.300 / 5.885 / 25.594 / 200.
The models of the streaming demo and the Unity client. They predict the
neutral SMPL-X body with 10 shape parameters that drives the Unity avatar;
training/configs/gru_smplx.yaml and lstm_smplx.yaml train models in this
format.
| Checkpoint | Architecture | Input window | Used by |
|---|---|---|---|
GRU/baseline.pt |
GRU | 60 fps, T = 40 | python -m streaming.live_demo --model gru |
LSTM/baseline.pt |
LSTM | 30 fps, T = 40 | Unity on-device (exported with streaming/export_onnx.py to flowavatar_lstm.onnx), python -m streaming.live_demo --model lstm |