Skip to content
circle-groupPublic

About

Official implementation of the paper PoseD-Flow: Versatile and Controlled Flow Matching Model of Human Pose

Resources

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

PoseD-Flow

Official implementation of the paper PoseD-Flow: Versatile and Controlled Flow Matching Model of Human Pose

Create env

conda env create -f environment.yml
conda activate poserfm

Training Model

AMASS

The processed AMASS dataset used for training the model can be downloaded from GDrive.

gdown 1TQi_wKxJU3TTcVko-oPlWvp8L12lNR7F
mkdir -p data/amass
unzip AMASS.zip -d data/amass && rm AMASS.zip

Body Models

We use SMPL and SMPL-X human models in our experiments. Download and save them in data/human_model_files.

Train model

python train.py

or, with custom args

python train.py --pose_datadir path/to/amass --use_wandb False

Pretrained model

The trained model and config can be downloaded from GDrive.

gdown 13DVC2dIEBAO2l1Wu1Zz7YbRXm3dl3Qrp
unzip ckpt.zip && rm ckpt.zip

Sample poses

python pose_generation.py --num_steps 100 --solver midpoint

Downstream tasks

Pose completion

python pose_completion.py --occ_type left_leg
python pose_completion.py --occ_type legs
python pose_completion.py --occ_type arms --num_steps 300
python pose_completion.py --occ_type trunk

Motion denoising

We use HumanEva split of AMASS and HPS dataset in our experiments. Split the datasets into 60s chunks using the script provided experiment/create_chunks.py, and save them in data/[HumanEva/HPS]_60frame.

python experiment/create_chunks.py --input-dir data/HumanEva --output-dir data/HumanEva_60frame  --seq-len 60

To run the experiments:

python motion_denoising.py --dataset AMASS
python motion_denoising.py --dataset AMASS --noise 0.1 --num_steps 400 --temp_weight 1e-1
python motion_denoising.py --dataset HPS --num_steps 400
python motion_denoising.py --dataset HPS --noise 0.1 --num_steps 400 --temp_weight 1e-1

Human Mesh Recovery

We use the EHF dataset from SMPL-X. Download the EHF dataset from SMPL-X and detect the 2d keypoints using ViTPose. The expected directory structure is:

data/ehf/
├── EHF/
│   ├── 01_align.ply
│   ├── 01_img.jpg
│   ├── 01_img.png
│   ├── 01_scan.obj
│   └── ...
└── EHF_keypoints/
    └── vitpose_keypoints/
        └── predictions/
            ├── 01_img.json
            ├── 02_img.json
            └── ...

To run the experiment:

python human_mesh_recovery.py 

We also support initialization from CLIFF. Follow the steps in their README and save the poses as data/ehf/CLIFF/smpl_poses_cliff.npz. To run the experiment:

python human_mesh_recovery.py --cliff

For convenience, we provide the keypoints and CLIFF poses used in our experiments in GDrive.

gdown 1WLbujgAyNGHyjOWThwr-p5H8uBCQ1176
unzip ehf_data.zip && rm ehf_data.zip

Citation

If you find this code helpful for your research, please cite our paper:

    @inproceedings{nadar2026posedflow,
        title={PoseD-Flow: Versatile and Guided Flow Matching Model of Human Pose},
        author={Jebastin Nadar and Simone Foti and Tolga Birdal},
        booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
        pages={21165--21175},
        year={2026}
    }

About

Official implementation of the paper PoseD-Flow: Versatile and Controlled Flow Matching Model of Human Pose

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages