Lingteng Qiu*, Xiaodong Gu*, Peihao Li*, Qi Zuo*, Weichao Shen, Junfei Zhang, Kejie Qiu, Weihao Yuan
Guanying Chen+, Zilong Dong+, Liefeng Bo
如果您熟悉中文，可以阅读中文版本的README
[April 16, 2025] We have released a memory-saving version of motion and LHM. Now you can run the entire pipeline on 14 GB GPUs.
[April 13, 2025] We have released LHM-MINI, which allows you to run LHM on 16 GB GPUs. 🔥🔥🔥
[April 10, 2025] We release the motion extraction node and animation infer node of LHM on ComfyUI. With a extracted offline motion, you can generate a 10s animation clip in 20s!!! Update your ComfyUI branch right now.🔥🔥🔥
[April 9, 2025] we build a detailed tutorial to guide users to install LHM-ComfyUI on Windows step by step!
[April 9, 2025] We release the video processing pipeline to create your training data LHM_Track!
For more details about the updates, see 👉 👉 👉 logger.
- Core Inference Pipeline (v0.1) 🔥🔥🔥
- HuggingFace Demo Integration 🤗🤗🤗
- ModelScope Deployment
- Motion Processing Scripts
- Training Codes Release
We provide a video that teaches us how to install LHM and LHM-ComfyUI step by step on YouTube, submitted by softicelee2.
We provide a video that teaches us how to install LHM step by step on bilibili, submitted by 站长推荐推荐.
We provide a video that teaches us how to install LHM-ComfyUI step by step on bilibili, submitted by 站长推荐推荐.
Please sure you had install nvidia-docker in our system.
# Linux System only
# CUDA 121
# step0. download docker images
wget -P ./lhm_cuda_dockers https://virutalbuy-public.oss-cn-hangzhou.aliyuncs.com/share/aigc3d/data/for_lingteng/LHM/LHM_Docker/lhm_cuda121.tar
# step1. build from docker file
sudo docker load -i ./lhm_cuda_dockers/lhm_cuda121.tar
# step2. run docker_file and open the communication port 7860
sudo docker run -p 7860:7860 -v PATH/FOLDER:DOCKER_WORKSPACES -it lhm:cuda_121 /bin/bash
Clone the repository.
git clone [email protected]:aigc3d/LHM.git
cd LHMSet Up a Virtual Environment Open Command Prompt (CMD), navigate to the project folder, and run:
python -m venv lhm_env
lhm_env\Scripts\activate
install_cu121.bat
python ./app.py# cuda 11.8
pip install rembg
sh ./install_cu118.sh
# cuda 12.1
sh ./install_cu121.shThe installation has been tested with python3.10, CUDA 11.8 or CUDA 12.1. Or you can install dependencies step by step, following INSTALL.md.
Please note that the model will be downloaded automatically if you do not download it yourself.
| Model | Training Data | BH-T Layers | ModelScope | HuggingFace | Inference Time | input requirement |
|---|---|---|---|---|---|---|
| LHM-MINI | 300K Videos + 5K Synthetic Data | 2 | ModelScope | huggingface | 1.41 s | half & full body |
| LHM-500M | 300K Videos + 5K Synthetic Data | 5 | ModelScope | huggingface | 2.01 s | full body |
| LHM-500M-HF | 300K Videos + 5K Synthetic Data | 5 | ModelScope | huggingface | 2.01 s | half & full body |
| LHM-1.0B | 300K Videos + 5K Synthetic Data | 15 | ModelScope | huggingface | 6.57 s | full body |
| LHM-1B-HF | 300K Videos + 5K Synthetic Data | 15 | ModelScope | huggingface | 6.57 s | half & full body |
Model cards with additional details can be found in model_card.md.
from huggingface_hub import snapshot_download
model_dir = snapshot_download(repo_id='3DAIGC/LHM-MINI', cache_dir='./pretrained_models/huggingface')
# 500M-HF Model
model_dir = snapshot_download(repo_id='3DAIGC/LHM-500M-HF', cache_dir='./pretrained_models/huggingface')
# 1B-HF Model
model_dir = snapshot_download(repo_id='3DAIGC/LHM-1B-HF', cache_dir='./pretrained_models/huggingface')from modelscope import snapshot_download
model_dir = snapshot_download(model_id='Damo_XR_Lab/LHM-MINI', cache_dir='./pretrained_models')
# 500M-HF Model
model_dir = snapshot_download(model_id='Damo_XR_Lab/LHM-500M-HF', cache_dir='./pretrained_models')
# 1B-HF Model
model_dir = snapshot_download(model_id='Damo_XR_Lab/LHM-1B-HF', cache_dir='./pretrained_models')# Download prior model weights
wget https://virutalbuy-public.oss-cn-hangzhou.aliyuncs.com/share/aigc3d/data/LHM/LHM_prior_model.tar
tar -xvf LHM_prior_model.tar We provide the test motion examples, we will update the processing scripts ASAP :).
# Download prior model weights
wget https://virutalbuy-public.oss-cn-hangzhou.aliyuncs.com/share/aigc3d/data/LHM/motion_video.tar
tar -xvf ./motion_video.tar After downloading weights and data, the folder of the project structure seems like:
├── configs
│ ├── inference
│ ├── accelerate-train-1gpu.yaml
│ ├── accelerate-train-deepspeed.yaml
│ ├── accelerate-train.yaml
│ └── infer-gradio.yaml
├── engine
│ ├── BiRefNet
│ ├── pose_estimation
│ ├── SegmentAPI
├── example_data
│ └── test_data
├── exps
│ ├── releases
├── LHM
│ ├── datasets
│ ├── losses
│ ├── models
│ ├── outputs
│ ├── runners
│ ├── utils
│ ├── launch.py
├── pretrained_models
│ ├── dense_sample_points
│ ├── gagatracker
│ ├── human_model_files
│ ├── sam2
│ ├── sapiens
│ ├── voxel_grid
│ ├── arcface_resnet18.pth
│ ├── BiRefNet-general-epoch_244.pth
├── scripts
│ ├── exp
│ ├── convert_hf.py
│ └── upload_hub.py
├── tools
│ ├── metrics
├── train_data
│ ├── example_imgs
│ ├── motion_video
├── inference.sh
├── README.md
├── requirements.txtNow, we support user motion sequence input. As the pose estimator requires some GPU memory, this Gradio application requires at least 24 GB of GPU memory to run LHM-500M.
# Memory-saving version; More time available for Use.
# The maximum supported length for 720P video is 20s.
python ./app_motion_ms.py
python ./app_motion_ms.py --model_name LHM-1B-HF
# Support user motion sequence input. As the pose estimator requires some GPU memory, this Gradio application requires at least 24 GB of GPU memory to run LHM-500M.
python ./app_motion.py
python ./app_motion.py --model_name LHM-1B-HF
# preprocessing video sequence
python ./app.py
python ./app.py --model_name LHM-1B
Now we support upper-body image input!
# MODEL_NAME={LHM-500M-HF, LHM-500M, LHM-1B, LHM-1B-HF}
# bash ./inference.sh LHM-500M-HF ./train_data/example_imgs/ ./train_data/motion_video/mimo1/smplx_params
# bash ./inference.sh LHM-500M ./train_data/example_imgs/ ./train_data/motion_video/mimo1/smplx_params
# bash ./inference.sh LHM-1B ./train_data/example_imgs/ ./train_data/motion_video/mimo1/smplx_params
# animation
bash inference.sh ${MODEL_NAME} ${IMAGE_PATH_OR_FOLDER} ${MOTION_SEQ}
# export mesh
bash ./inference_mesh.sh ${MODEL_NAME} -
Download model weights for motion processing.
wget -P ./pretrained_models/human_model_files/pose_estimate https://virutalbuy-public.oss-cn-hangzhou.aliyuncs.com/share/aigc3d/data/LHM/yolov8x.pt wget -P ./pretrained_models/human_model_files/pose_estimate https://virutalbuy-public.oss-cn-hangzhou.aliyuncs.com/share/aigc3d/data/LHM/vitpose-h-wholebody.pth
-
Install extra dependencies.
cd ./engine/pose_estimation pip install mmcv==1.3.9 pip install -v -e third-party/ViTPose pip install ultralytics -
Run the script.
# python ./engine/pose_estimation/video2motion.py --video_path ./train_data/demo.mp4 --output_path ./train_data/custom_motion python ./engine/pose_estimation/video2motion.py --video_path ${VIDEO_PATH} --output_path ${OUTPUT_PATH} # for half-body video, e.g. ./train_data/xiaoming.mp4, we recommend to use command as below: python ./engine/pose_estimation/video2motion.py --video_path ${VIDEO_PATH} --output_path ${OUTPUT_PATH} --fitting_steps 100 0
-
Use the motion to drive the avatar.
# if not sam2? pip install rembg. # bash ./inference.sh LHM-500M-HF ./train_data/example_imgs/ ./train_data/custom_motion/demo/smplx_params # bash ./inference.sh LHM-1B-HF ./train_data/example_imgs/ ./train_data/custom_motion/demo/smplx_params bash inference.sh ${MODEL_NAME} ${IMAGE_PATH_OR_FOLDER} ${OUTPUT_PATH}/${VIDEO_NAME}/smplx_params
We provide some simple scripts to compute the metrics.
# download pretrain model into ./pretrained_models/
wget https://virutalbuy-public.oss-cn-hangzhou.aliyuncs.com/share/aigc3d/data/LHM/arcface_resnet18.pth
# Face Similarity
python ./tools/metrics/compute_facesimilarity.py -f1 ${gt_folder} -f2 ${results_folder}
# PSNR
python ./tools/metrics/compute_psnr.py -f1 ${gt_folder} -f2 ${results_folder}
# SSIM LPIPS
python ./tools/metrics/compute_ssim_lpips.py -f1 ${gt_folder} -f2 ${results_folder} We have implemented a standard workflow and related nodes for customlize video animation. You can use any character and any driven videos this time! See branch feat/comfyui for more information!
We need a comfyui windows install guide of our feat/comfyui branch. If you are familiar with comfyui and successfully install it on windows, welcome to submit a pr to update windows install guide for our community!
This work is built on many amazing research works and open-source projects:
Thanks for their excellent works and great contribution to 3D generation and 3D digital human area.
We would like to express our sincere gratitude to 站长推荐推荐 and softicelee2 for the installation tutorial video on bilibili.
Welcome to follow our team other interesting works:
@inproceedings{qiu2025LHM,
title={LHM: Large Animatable Human Reconstruction Model from a Single Image in Seconds},
author={Lingteng Qiu and Xiaodong Gu and Peihao Li and Qi Zuo
and Weichao Shen and Junfei Zhang and Kejie Qiu and Weihao Yuan
and Guanying Chen and Zilong Dong and Liefeng Bo
},
booktitle={arXiv preprint arXiv:2503.10625},
year={2025}
}
