ICCV, 2025
Shaowei Liu · Chuan Guo* · Bing Zhou* · Jian Wang* ·

This repository contains the pytorch implementation for the paper Ponimator: Unfolding Interactive Pose for Versatile Human-Human Interaction Animation, ICCV 2025. In this paper, we propose a uniform framework for human-human interaction animation and generation anchored on interactive poses.

• Installation
• Interactive Pose Animation Demo
• Interactive Motion Generation Demo
• Training and Inference
• Custom Third-party Scripts
• Citation

• Clone this repository:

  git clone https://github.com/stevenlsw/ponimator.git
  cd ponimator

• Our code has minimal dependencies:

  # most version of python and pytorch should all work
  conda create -n ponimator python=3.10
  conda activate ponimator
  pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu124
  pip3 install aitviewer yacs roma git+https://github.com/openai/CLIP.git

• [Optional] Install visualization dependencies:

  pip install blendify --find-links https://download.blender.org/pypi/bpy/
  pip install imageio[ffmpeg]

• Prepare SMPL-X body models, you can download a minimal version from minmal smplx folder and put smplx folder under body_models/

  body_models/ 
    ├── smplx
      ├── SMPLX_MALE.npz
      ├── SMPLX_FEMALE.npz
      ├── SMPLX_NEUTRAL.npz
      ├── SMPLX_MALE.pkl
      ├── SMPLX_FEMALE.pkl
      ├── SMPLX_NEUTRAL.pkl

• Download pretrained checkpoints: Interactive Pose Animator and Interactive Pose Generator and put under checkpoints/

  checkpoints/ 
  ├── contactmotion.ckpt
  ├── contactpose.ckpt

• [Optional] Given an interactive pose image, you can estimate the interactive pose by buddi check custom third-party scripts

• We put buddi demo outputs under data/buddi/{video_name}, you can generate interactive motion by:

  python scripts/run_pose2motion.py --data_dir data/buddi/Couple_6806 --save

• Visualize the interactive pose by append --vis_interactive_pose to above command. For off-screen rendering, you can add xvfb-run -a -s "-screen 0 1024x768x24" before the command. The output video will be saved under outputs/Couple_6806. You can also turn off visualization by append --disable_vis to the command. The output video will be saved under outputs/Couple_6806/vis_motion_pred.mp4.

• [Optional] To render the generated motion on original image:

  python demo/vis_smpler.py --data_dir data/buddi/Couple_6806 --result_dir outputs/Couple_6806

  The output video will be saved under outputs/Couple_6806/render_motion_pred.mp4.

• You are expected to get the following result:

  Input Image	Generated MOtion	Output Video

• [Optional] Given an single-person image, you can use SMPLer-X to estimate the single pose or animate existing single-person pose dataset like Motion-X. See custom third-party scripts for more details.

• We put demo Motion-X data under data/motionx/{video_name}, you can generate interactive motion by:

  python scripts/run_singlepose2motion.py --data_dir data/motionx/Back_Flip_Kungfu_wushu_Trim9_clip1 --save

  The output video will be saved under outputs/Back_Flip_Kungfu_wushu_Trim9_clip1/vis_motion_pred.mp4. You can adjust --seed, --inter_time_idx, --gender, --text for different outputs.

• [Optional] To render the generated motion on original image:

  python demo/vis_motionx.py --data_dir data/motionx/Back_Flip_Kungfu_wushu_Trim9_clip1 --result_dir outputs/Back_Flip_Kungfu_wushu_Trim9_clip1

  The output video will be saved under outputs/Back_Flip_Kungfu_wushu_Trim9_clip1/render_motion_pred.mp4.

• You are expected to get the following result:

  Input Image	Generated MOtion	Output Video

• Estimate Interactive Pose by Buddi, a custom script is at third_party_scripts/buddi/custom_demo.sh, put under same directory as buddi root dir.

  cd buddi/
  chmod +x /custom_demo.sh
  ./custom_demo.sh ${input_image_path} {gpu_id}

• Estimate Single Pose by SMPLest-X, a custom script is at third_party_scripts/SMPLest-X/custom_inference.py, put under same directory as SMPLest-X/main folder.

  cd SMPLest-X
  python main/custom_inference.py --file_name ${input_image_path} --ckpt_name smplest_x_h --output_folder ${output_folder} --save_pkl

• Example estimated single pose from SMPLest-X is shown as

  Input Image	Estimated Single Pose

  • Quick Note: The example image is the most famous statue in Waikiki beach, Hawaii.

• Generate interactive motion of custom text "two person hug each other" by

  python scripts/run_singlepose2motion.py --save  --data_dir data/smpler --data_source smpler  --text "two person hug each other" --seed 1 --save_dir outputs/smpler

  The default output is under outputs/image_ori.

• Render the output video on original image:

  python demo/vis_smpler.py --data_dir data/smpler --result_dir outputs/image_ori

  The output video will be saved under outputs/image_ori/rendered_video.mp4.

• You are expected to get the following result:

  Input Image	Generated MOtion	Output Video

If you find our work useful in your research, please cite:

@inproceedings{liu2025ponimator},
  title={Ponimator: Unfolding Interactive Pose for Versatile Human-Human Interaction Animation},
  author={Liu, Shaowei and Guo, Chuan and Zhou, Bing and Wang, Jian},
  booktitle={International Conference on Computer Vision (ICCV)},
  year={2025}
}

Our model is based on InterGen . We are also grateful to several other open-source repositories that we built upon during the development of our pipeline:

• Buddi for interactive pose estimation.
• SMPLer-X for single pose estimation.
• AitViewer for visualization.