Human3R: Everyone Everywhere All at Once

Yue Chen, Xingyu Chen*, Yuxuan Xue, Anpei Chen, Yuliang Xiu†, Gerard Pons-Moll

TL;DR: Inference with One model, One stage; Training in One day using One GPU

human3r.mp4

Getting Started

Installation

1. Clone Human3R.

git clone https://github.com/fanegg/Human3R.git
cd Human3R

2. Create the environment.

conda create -n human3r python=3.11 cmake
conda activate human3r

# PyTorch — pick the build that matches your GPU.
# Most setups (CUDA 12.4):
conda install pytorch torchvision pytorch-cuda=12.4 -c pytorch -c nvidia
# Blackwell GPUs (RTX 5060/5070/5080/5090) need CUDA 12.8 + nightly torch instead:
# pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128

pip install -r requirements.txt
# issues with pytorch dataloader, see https://github.com/pytorch/pytorch/issues/99625
conda install 'llvm-openmp<16'
# compilers + ninja for building the cuda kernels below (gsplat, pytorch3d, RoPE)
conda install -y gcc_linux-64 gxx_linux-64
pip install ninja

pip install git+https://github.com/nerfstudio-project/gsplat.git

# VolumetricSMPL + pytorch3d (used by demo.py for the volumetric body model / penetration eval).
# pytorch3d has no prebuilt wheels for recent torch/CUDA combos, so build from source:
pip install fvcore iopath
pip install --no-build-isolation "git+https://github.com/facebookresearch/pytorch3d.git@main"
pip install VolumetricSMPL

# for evaluation
pip install evo
pip install open3d

3. Compile the cuda kernels for RoPE (as in CroCo v2).

cd src/croco/models/curope/
python setup.py build_ext --inplace
cd ../../../../

Download

Run the following commands to download all models and checkpoints into the src/ directory. The first command will prompt you to register and log in to access each version of SMPL.

# SMPLX family models
bash scripts/fetch_smplx.sh

# Human3R checkpoints
huggingface-cli download faneggg/human3r human3r_896L.pth --local-dir ./src

Inference Demo

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

# input can be a folder or a video
# the following script will run inference with Human3R and visualize the output with viser on port 8080
CUDA_VISIBLE_DEVICES=0 python demo.py --model_path MODEL_PATH --size 512 \
    --seq_path SEQ_PATH --output_dir OUT_DIR --subsample 1 --use_ttt3r \
    --vis_threshold 2 --downsample_factor 1 --reset_interval 100

# Example:
# To save the results, append `--save --output_dir tmp` to the command.
CUDA_VISIBLE_DEVICES=0 python demo.py --model_path src/human3r_896L.pth \
    --size 512 --seq_path examples/GoodMornin1.mp4 \
    --subsample 1 --use_ttt3r --vis_threshold 2 \
    --downsample_factor 1 --reset_interval 100

Evaluation

Please refer to the eval.md for more details.

Model Cards

Please refer to the inference.md for using different backbones.

Training

Please refer to the train.md for more details.

Acknowledgements

Our code is based on the following awesome repositories:

• CUT3R, TTT3R, Multi-HMR, PromptHMR, GVHMR, MonST3R, Easi3R, DUSt3R, Viser, BEDLAM

We thank the authors for releasing their code!

Citation

If you find our work useful, please cite:

@article{chen2025human3r,
    title={Human3R: Everyone Everywhere All at Once},
    author={Chen, Yue and Chen, Xingyu and Xue, Yuxuan and Chen, Anpei and Xiu, Yuliang and Gerard, Pons-Moll},
    journal={arXiv preprint arXiv:2510.06219},
    year={2025}
    }