BLADE: Single-view Body Mesh Estimation through Accurate Depth Estimation (CVPR 2025)

Authors: Shengze Wang, Jiefeng Li, Tianye Li, Ye Yuan, Henry Fuchs, Koki Nagano*, Shalini De Mello*, Michael Stengel*
*denotes equal contribution

Overview

Unlike prior methods that avoid camera recovery and only work well for subjects far from the camera, BLADE tackles close-range human mesh recovery where perspective distortion is strongest in 3 steps:

Predicts pelvis depth (Tz): directly from the image
Conditions pose/shape on Tz: the same shape/pose can look different at different distances
Recovers full perspective camera: focal length and XY-translation through differentiable rasterization

Key Features

✅ True perspective HMR: recovers focal length and full 3D translation without heuristics
✅ Close-range robust: strong performance under severe perspective distortion
✅ 3D Accurate & 2D Aligned: improved 3D pose, metrical depth, and 2D reprojection alignment

📚 Citation & Acknowledgement

@inproceedings{wang2025blade,
  title     = {BLADE: Single-view Body Mesh Estimation through Accurate Depth Estimation},
  author    = {Wang, Shengze and Li, Jiefeng and Li, Tianye and Yuan, Ye and Fuchs, Henry and Nagano, Koki and De Mello, Shalini and Stengel, Michael},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2025}
}

BLADE builds on the outstanding work and open-source contributions of many teams. We thank:

MMHuman3D — dataset preprocessing utilities, evaluation protocols, and loaders that informed our data pipeline.
ZOLLY & PDHuman — PDHuman dataset and related preprocessing guidance and ZOLLY as baseline.
SMPL & SMPL-X body models (MPI-IS) — foundational parametric human body models.
Depth Anything v2 (metric) — depth encoder's features are used in our DepthNet.
MediaPipe & Sapiens Pose & RTMDet — human segmenter and keypoint models used in our preprocessing pipeline.
AiOS & ControlNet — basis for our pelvis-depth conditioned PoseNet.
Datasets — We gratefully acknowledge the following datasets used for training/evaluation and comparisons: HuMMan, Human3.6M, PDHuman, SPEC, BEDLAM (basis for our BEDLAM-CC renders)

If you use BLADE in your research, please also cite the original works for any datasets/models you use alongside our paper.

For business inquiries, please visit our website and submit the form: NVIDIA Research Licensing.

🛠️ Installation

Prerequisites

CUDA 11.8 recommended. guide for CUDA 13 also provided
Python 3.9.19
Conda

Quick Installation (CUDA 11.8)

# Create environment
apt-get install git-lfs
git lfs install

git clone https://github.com/NVlabs/BLADE

conda create -n blade_env python=3.9.19
conda deactivate && conda activate blade_env

# pytorch stuff
pip install tqdm torch==2.0.1+cu118 torchvision==0.15.2+cu118 --extra-index-url https://download.pytorch.org/whl/cu118 && pip install fvcore iopath numpy==1.24.4 && pip install wandb && pip install --no-index --no-cache-dir pytorch3d -f https://dl.fbaipublicfiles.com/pytorch3d/packaging/wheels/py39_cu118_pyt201/download.html && pip install --no-warn-conflict matplotlib==3.8.4 colorama requests huggingface-hub safetensors pillow six click openxlab && pip install --no-warn-conflict chumpy scipy munkres tqdm cython fsspec yapf==0.40.1 packaging omegaconf ipdb ftfy regex && pip install --no-warn-conflict json_tricks terminaltables modelindex prettytable albumentations && pip install --no-warn-conflict smplx==0.1.28 debugpy numba yacs scikit-learn filterpy h5py trimesh scikit-image tensorboardx pyrender torchgeometry joblib boto3 easydict pycocotools colormap pytorch-transformers pickle5 plyfile timm pyglet future tensorboard cdflib ftfy einops && pip install --no-warn-conflict numpy==1.23.1 mediapipe
# MMCV stuff
cd mmcv; MMCV_WITH_OPS=1 pip install --no-warn-conflicts -e . -v; pip install --no-warn-conflicts -e .; cd ..; cd sapiens; cd engine; pip install --no-warn-conflicts -e . -v; cd ../pretrain; pip install --no-warn-conflicts -e . -v; cd ../pose; pip install --no-warn-conflicts -e . -v; cd ../det; pip install --no-warn-conflicts -e . -v; cd ../seg; pip install --no-warn-conflicts -e . -v; cd ../..; pip install --no-warn-conflicts -e .;
pip install --no-warn-conflict ffmpeg astropy easydev pandas rtree vedo codecov flake8 interrogate isort pytest surrogate xdoctest setuptools loguru open3d omegaconf
cd aios_repo/models/aios/ops/; python setup.py build install; cd ../../../..; cd torch-trust-ncg; python setup.py install; 
cd ..; pip install  --no-warn-conflict numpy==1.23.1

Note: Ignore numpy version warnings - they don't affect functionality.

For non-CUDA 11.8 Users (e.g. CUDA 12 & CUDA 13)

If you don't want to install another cuda, try changing the cu118 in the above commands to your version. This sometimes works, depending on whether pytorch and pytorch3d has a matching version for you.

The more reliable way would be following this guide (requires installing CUDA 11.8) INSTALL_CUDA13.md for detailed non-CUDA 11.8 instructions.

FYI a system could have multiple versions of CUDA at the same time.

📦 Download Models

Required Models

# Install HuggingFace CLI
pip install "huggingface_hub[cli]"
huggingface-cli login
# Note: You might need to set up your HuggingFace token in order to download

# Download BLADE checkpoint
Pretrained weights are personally provided by Shengze Wang 
Please check https://mcmvmc.github.io/blade_weights.html 

# If you are on SLURM and limited by disk quota, redirect the download directories
export HF_HOME=</path/to/.cache/>; 
export TORCH_HOME=</path/to/.cache/>; 
export WANDB_DIR=</path/to/.wandb_cache/>; 

# Download supporting models
hf download depth-anything/Depth-Anything-V2-Metric-Hypersim-Large depth_anything_v2_metric_hypersim_vitl.pth --local-dir pretrained/model_init_weights
hf download ttxskk/AiOS aios_checkpoint.pth --local-dir pretrained/model_init_weights
hf download facebook/sapiens-pose-bbox-detector rtmdet_m_8xb32-100e_coco-obj365-person-235e8209.pth --local-dir pretrained/rtmpose
hf download facebook/sapiens-pose-1b sapiens_1b_goliath_best_goliath_AP_639.pth --local-dir pretrained/pose

SMPL/SMPL-X Models (Manual Download Required)

Due to licensing restrictions, download these manually:

SMPL: Download from official site → place in body_models/smpl/
SMPL-X: Download from official site → place in body_models/smplx/
Transfer data: Download from SMPL-X repo → place in pretrained/transfer_data/

Expected Directory Structure

blade_repo/
├── body_models/
│   ├── smpl/          # SMPL models (manual download)
│   └── smplx/         # SMPL-X models (manual download)
├── pretrained/
│   ├── epoch_2.pth    # BLADE checkpoint
│   ├── model_init_weights/
│   │   ├── depth_anything_v2_metric_hypersim_vitl.pth
│   │   └── aios_checkpoint.pth
│   ├── rtmpose/
│   │   └── rtmdet_m_8xb32-100e_coco-obj365-person-235e8209.pth
│   ├── pose/
│   │   └── sapiens_1b_goliath_best_goliath_AP_639.pth
│   └── transfer_data/ # SMPL conversion data (manual download)
│       ├── smplh2smpl_def_transfer.pkl
│       ├── smplx2smplh_deftrafo_setup.pkl
│       ├── smpl2smplh_def_transfer.pkl
│       ├── smplh2smplx_deftrafo_setup.pkl
│       ├── smplx_mask_ids.npy
│       ├── smpl2smplx_deftrafo_setup.pkl
│       ├── smplx2smpl_deftrafo_setup.pkl
│       └── smplx_to_smpl.pkl

🚀 Usage

Quick Demo

Although we can't share the photos used in our paper due to license limitations, we provide some CC0-licensed images for users to test quickly.

For in-the-wild demos, we enable pose & depth optimization by default (opt_pose=True and opt_tz=True in blade/configs/blade_inthewild.py) If you are running numeric evaluation, please turn off pose and depth optimization because the optimization could fail due to various factors such as bad 2D keypoint detections or segmentations.

MINI_BATCHSIZE=5 python api/test_api.py ./demo_images/

Note: If you see issue related to EGL or XRender (often on monitor-less servers like SLURM), do:

export PYGLET_HEADLESS=True
export PYOPENGL_PLATFORM=egl
export LD_LIBRARY_PATH="$CONDA_PREFIX/lib:${LD_LIBRARY_PATH}"
[ -f "$CONDA_PREFIX/lib/libEGL.so" ] || ln -s "$CONDA_PREFIX/lib/libEGL.so.1" "$CONDA_PREFIX/lib/libEGL.so"

Samples You Should See:

Results are written to results/test_demo by default. Yo

Blade

Install / Use

README