Learnable SMPLify: A Neural Solution for Optimization-Free Human Pose Inverse Kinematics

arXiv Paper | Hugging Face

TL;DR Given X_{t-s} and X_{t} 3D keypoints, calculate residual SMPL parameters from t-s to t.

Preparation

Refer to PREPARATION.md for installation and data preparation details.

Checkpoints

The pretrained model checkpoint is available at Google Drive and Hugging Face.

Usage

Training

cd to src folder and run the following command.

torchrun --nproc-per-node <NUM_GPUS> main.py --config configs/net.yaml (--extra_tag <EXTRA_TAG> --batch_size <BATCH_SIZE> --epochs <EPOCHS>)

You can get logs, tensorboard and checkpoints in the corresponding logs/<MODEL_NAME>_net_<EXTRA_TAG> folder.

Evaluation

To evaluate the model, run the following command:

torchrun --nproc-per-node <NUM_GPUS> main.py --config configs/net.yaml --eval --checkpoint <PATH_TO_CHECKPOINT>

Sequential Inference

To run sequential inference, you can use the following command:

python inference.py <PATH_TO_CHECKPOINT> (<DATASET_NAME> <SAMPLE_RATIO>)

Citation

If you find this work useful in your research, please consider citing:

@misc{LearnableSMPLify,
      title={Learnable SMPLify: A Neural Solution for Optimization-Free Human Pose Inverse Kinematics},
      author={Yuchen, Yang and Linfeng, Dong and Wei, Wang and Zhihang, Zhong and Xiao, Sun},
      year={2025},
      eprint={2508.13562},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Acknowledgement

We thank the authors of ST-GCN, ReFit, OSX for their great works. We partially refer to their codebases for this project.