[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-72434":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":16,"subscribersCount":16,"size":16,"stars1d":17,"stars7d":18,"stars30d":19,"stars90d":16,"forks30d":16,"starsTrendScore":18,"compositeScore":20,"rankGlobal":10,"rankLanguage":10,"license":21,"archived":22,"fork":22,"defaultBranch":23,"hasWiki":22,"hasPages":22,"topics":24,"createdAt":10,"pushedAt":10,"updatedAt":28,"readmeContent":29,"aiSummary":30,"trendingCount":16,"starSnapshotCount":16,"syncStatus":17,"lastSyncTime":31,"discoverSource":32},72434,"LHM","aigc3d\u002FLHM","aigc3d","[ICCV2025] LHM: Large Animatable Human Reconstruction Model from a Single Image in Seconds","https:\u002F\u002Faigc3d.github.io\u002Fprojects\u002FLHM\u002F",null,"Python",2626,209,40,46,0,2,6,18,67.77,"Apache License 2.0",false,"main",[25,26,27],"aicg","aigc","digitalhuman","2026-06-12 04:01:05","# \u003Cspan>\u003Cimg src=\".\u002Fassets\u002FLHM_logo_parsing.png\" height=\"35\" style=\"vertical-align: top;\"> - Official PyTorch Implementation\u003C\u002Fspan>\n\n#####  \u003Cp align=\"center\"> [Lingteng Qiu\u003Csup>*\u003C\u002Fsup>](https:\u002F\u002Flingtengqiu.github.io\u002F), [Xiaodong Gu\u003Csup>*\u003C\u002Fsup>](https:\u002F\u002Fscholar.google.com.hk\u002Fcitations?user=aJPO514AAAAJ&hl=zh-CN&oi=ao), [Peihao Li\u003Csup>*\u003C\u002Fsup>](https:\u002F\u002Fliphao99.github.io\u002F), [Qi Zuo\u003Csup>*\u003C\u002Fsup>](https:\u002F\u002Fscholar.google.com\u002Fcitations?user=UDnHe2IAAAAJ&hl=zh-CN), [Weichao Shen](https:\u002F\u002Fscholar.google.com\u002Fcitations?user=7gTmYHkAAAAJ&hl=zh-CN), [Junfei Zhang](https:\u002F\u002Fscholar.google.com\u002Fcitations?user=oJjasIEAAAAJ&hl=en), [Kejie Qiu](https:\u002F\u002Fsites.google.com\u002Fsite\u002Fkejieqiujack\u002Fhome), [Weihao Yuan](https:\u002F\u002Fweihao-yuan.com\u002F)\u003Cbr> [Guanying Chen\u003Csup>+\u003C\u002Fsup>](https:\u002F\u002Fguanyingc.github.io\u002F), [Zilong Dong\u003Csup>+\u003C\u002Fsup>](https:\u002F\u002Fbaike.baidu.com\u002Fitem\u002F%E8%91%A3%E5%AD%90%E9%BE%99\u002F62931048), [Liefeng Bo](https:\u002F\u002Fscholar.google.com\u002Fcitations?user=FJwtMf0AAAAJ&hl=zh-CN)\u003C\u002Fp>\n#####  \u003Cp align=\"center\"> Tongyi Lab, Alibaba Group \u003Cbr> ICCV 2025 \u003C\u002Fp>\n\n[![Project Website](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F🌐-Project_Website-blueviolet)](https:\u002F\u002Faigc3d.github.io\u002Fprojects\u002FLHM\u002F)\n[![arXiv Paper](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F📜-arXiv:2503-10625)](https:\u002F\u002Farxiv.org\u002Fpdf\u002F2503.10625)\n[![HuggingFace](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F🤗-HuggingFace_Space-blue)](https:\u002F\u002Fhuggingface.co\u002Fspaces\u002FDyrusQZ\u002FLHM)\n[![ModelScope](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F%20ModelScope%20-Space-blue)](https:\u002F\u002Fwww.modelscope.cn\u002Fstudios\u002FDamo_XR_Lab\u002FLHM) \n[![MotionShop2](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F%20MotionShop2%20-Space-blue)](https:\u002F\u002Fmodelscope.cn\u002Fstudios\u002FDamo_XR_Lab\u002FMotionshop2) \n[![Apache License](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F📃-Apache--2.0-929292)](https:\u002F\u002Fwww.apache.org\u002Flicenses\u002FLICENSE-2.0)\n\n\n\u003Cp align=\"center\">\n  \u003Cimg src=\".\u002Fassets\u002FLHM_teaser.png\" heihgt=\"100%\">\n\u003C\u002Fp>\n\n如果您熟悉中文，可以[阅读中文版本的README](.\u002FREADME_CN.md)\n## 📢 Latest Updates\n**[March 2026]** **LHM++ is now open-sourced!** Supports arbitrary view inputs with higher efficiency—8-view input runs on just 8GB GPU memory—and superior rendering quality. See [GitHub](https:\u002F\u002Fgithub.com\u002Faigc3d\u002FLHM-plusplus) | [arXiv](https:\u002F\u002Farxiv.org\u002Fabs\u002F2506.13766) \u003Cbr>\n**[June 26, 2025]** LHM is got accepted by ICCV2025!!! \u003Cbr>\n**[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. \u003Cbr>\n**[April 13, 2025]** We have released LHM-MINI, which allows you to run LHM on 16 GB GPUs. 🔥🔥🔥 \u003Cbr>\n**[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](https:\u002F\u002Fgithub.com\u002Faigc3d\u002FLHM\u002Ftree\u002Ffeat\u002Fcomfyui) branch right now.🔥🔥🔥 \n\u003Cbr>\n**[April 9, 2025]** we build a detailed tutorial to guide users to install [LHM-ComfyUI](https:\u002F\u002Fgithub.com\u002Faigc3d\u002FLHM\u002Fblob\u002Ffeat\u002Fcomfyui\u002FWindows11_install.md) on Windows step by step!\u003Cbr>\n**[April 9, 2025]** We release the video processing pipeline to create your training data [LHM_Track](https:\u002F\u002Fgithub.com\u002Faigc3d\u002FLHM_Track)!\u003Cbr>\n\nFor more details about the updates, see 👉 👉 👉 [logger](.\u002Fassets\u002FNews_logger.md).\n\n### TODO List \n- [x] Core Inference Pipeline (v0.1) 🔥🔥🔥\n- [x] HuggingFace Demo Integration 🤗🤗🤗\n- [x] ModelScope Deployment\n- [x] Motion Processing Scripts\n- [ ] Release Training data & Testing Data (License Available) \n- [ ] Training Codes Release\n\n## 🚀 Getting Started\n\n\nWe provide a [video](https:\u002F\u002Fyoutu.be\u002FQ56Jllz33tk) that teaches us how to install LHM and LHM-ComfyUI step by step on YouTube, submitted by [softicelee2](https:\u002F\u002Fgithub.com\u002Fsofticelee2).\n\nWe provide a [video](https:\u002F\u002Fwww.bilibili.com\u002Fvideo\u002FBV18So4YCESk\u002F) that teaches us how to install LHM step by step on bilibili, submitted by 站长推荐推荐.\n\nWe provide a [video](https:\u002F\u002Fwww.bilibili.com\u002Fvideo\u002FBV1J9Z1Y2EiJ\u002F) that teaches us how to install LHM-ComfyUI step by step on bilibili, submitted by 站长推荐推荐.\n\n\n\n\n### Build from Docker\nPlease sure you had install nvidia-docker in our system.\n```\n# Linux System only\n# CUDA 121\n# step0. download docker images\nwget -P .\u002Flhm_cuda_dockers https:\u002F\u002Fvirutalbuy-public.oss-cn-hangzhou.aliyuncs.com\u002Fshare\u002Faigc3d\u002Fdata\u002Ffor_lingteng\u002FLHM\u002FLHM_Docker\u002Flhm_cuda121.tar \n\n# step1. build from docker file\nsudo docker load -i  .\u002Flhm_cuda_dockers\u002Flhm_cuda121.tar \n\n# step2. run docker_file and open the communication port 7860\nsudo docker run -p 7860:7860 -v PATH\u002FFOLDER:DOCKER_WORKSPACES -it lhm:cuda_121 \u002Fbin\u002Fbash\n```\n\n\n\n### Environment Setup\nClone the repository.\n```bash\ngit clone git@github.com:aigc3d\u002FLHM.git\ncd LHM\n```\n### Windows Installation\nSet Up a Virtual Environment\nOpen **Command Prompt (CMD)**, navigate to the project folder, and run:  \n```bash\npython -m venv lhm_env\nlhm_env\\Scripts\\activate\ninstall_cu121.bat\n\npython .\u002Fapp.py\n```\n\n```bash\n# cuda 11.8\npip install rembg\nsh .\u002Finstall_cu118.sh\n\n# cuda 12.1\nsh .\u002Finstall_cu121.sh\n```\nThe installation has been tested with python3.10, CUDA 11.8 or CUDA 12.1.\nOr you can install dependencies step by step, following [INSTALL.md](INSTALL.md).\n\n### Model Weights \n\n\u003Cspan style=\"color:red\">Please note that the model will be downloaded automatically if you do not download it yourself.\u003C\u002Fspan>\n\n| Model | Training Data | BH-T Layers | ModelScope| HuggingFace |Inference Time|input requirement|\n| :--- | :--- | :--- | :--- | :--- | :--- |:--- |\n| LHM-MINI | 300K Videos + 5K Synthetic Data | 2 | [ModelScope](https:\u002F\u002Fmodelscope.cn\u002Fmodels\u002FDamo_XR_Lab\u002FLHM-MINI) |[huggingface](https:\u002F\u002Fhuggingface.co\u002F3DAIGC\u002FLHM-MINI)| 1.41 s | half & full body|\n| LHM-500M | 300K Videos + 5K Synthetic Data | 5 | [ModelScope](https:\u002F\u002Fmodelscope.cn\u002Fmodels\u002FDamo_XR_Lab\u002FLHM-500M) |[huggingface](https:\u002F\u002Fhuggingface.co\u002F3DAIGC\u002FLHM-500M)| 2.01 s | full body|\n| LHM-500M-HF | 300K Videos + 5K Synthetic Data | 5 | [ModelScope](https:\u002F\u002Fmodelscope.cn\u002Fmodels\u002FDamo_XR_Lab\u002FLHM-500M-HF) |[huggingface](https:\u002F\u002Fhuggingface.co\u002F3DAIGC\u002FLHM-500M-HF)| 2.01 s | half & full body|\n| LHM-1.0B | 300K Videos + 5K Synthetic Data | 15 | [ModelScope](https:\u002F\u002Fmodelscope.cn\u002Fmodels\u002FDamo_XR_Lab\u002FLHM-1B) |[huggingface](https:\u002F\u002Fhuggingface.co\u002F3DAIGC\u002FLHM-1B)| 6.57 s | full body|\n| LHM-1B-HF | 300K Videos + 5K Synthetic Data | 15 | [ModelScope](https:\u002F\u002Fmodelscope.cn\u002Fmodels\u002FDamo_XR_Lab\u002FLHM-1B-HF) |[huggingface](https:\u002F\u002Fhuggingface.co\u002F3DAIGC\u002FLHM-1B-HF)| 6.57 s | half & full body|\n\n\nModel cards with additional details can be found in [model_card.md](modelcard.md).\n\n\n#### Download from HuggingFace\n```python\nfrom huggingface_hub import snapshot_download \nmodel_dir = snapshot_download(repo_id='3DAIGC\u002FLHM-MINI', cache_dir='.\u002Fpretrained_models\u002Fhuggingface')\n# 500M-HF Model\nmodel_dir = snapshot_download(repo_id='3DAIGC\u002FLHM-500M-HF', cache_dir='.\u002Fpretrained_models\u002Fhuggingface')\n# 1B-HF Model\nmodel_dir = snapshot_download(repo_id='3DAIGC\u002FLHM-1B-HF', cache_dir='.\u002Fpretrained_models\u002Fhuggingface')\n```\n\n#### Download from ModelScope \n```python\n\nfrom modelscope import snapshot_download\nmodel_dir = snapshot_download(model_id='Damo_XR_Lab\u002FLHM-MINI', cache_dir='.\u002Fpretrained_models')\n# 500M-HF Model\nmodel_dir = snapshot_download(model_id='Damo_XR_Lab\u002FLHM-500M-HF', cache_dir='.\u002Fpretrained_models')\n# 1B-HF Model\nmodel_dir = snapshot_download(model_id='Damo_XR_Lab\u002FLHM-1B-HF', cache_dir='.\u002Fpretrained_models')\n```\n\n### Download Prior Model Weights \n```bash\n# Download prior model weights\nwget https:\u002F\u002Fvirutalbuy-public.oss-cn-hangzhou.aliyuncs.com\u002Fshare\u002Faigc3d\u002Fdata\u002FLHM\u002FLHM_prior_model.tar \ntar -xvf LHM_prior_model.tar \n```\n\n### Data Motion Preparation\nWe provide the test motion examples, we will update the processing scripts ASAP :).\n\n```bash\n# Download prior model weights\nwget https:\u002F\u002Fvirutalbuy-public.oss-cn-hangzhou.aliyuncs.com\u002Fshare\u002Faigc3d\u002Fdata\u002FLHM\u002Fmotion_video.tar\ntar -xvf .\u002Fmotion_video.tar \n```\n\nAfter downloading weights and data, the folder of the project structure seems like:\n```bash\n├── configs\n│   ├── inference\n│   ├── accelerate-train-1gpu.yaml\n│   ├── accelerate-train-deepspeed.yaml\n│   ├── accelerate-train.yaml\n│   └── infer-gradio.yaml\n├── engine\n│   ├── BiRefNet\n│   ├── pose_estimation\n│   ├── SegmentAPI\n├── example_data\n│   └── test_data\n├── exps\n│   ├── releases\n├── LHM\n│   ├── datasets\n│   ├── losses\n│   ├── models\n│   ├── outputs\n│   ├── runners\n│   ├── utils\n│   ├── launch.py\n├── pretrained_models\n│   ├── dense_sample_points\n│   ├── gagatracker\n│   ├── human_model_files\n│   ├── sam2\n│   ├── sapiens\n│   ├── voxel_grid\n│   ├── arcface_resnet18.pth\n│   ├── BiRefNet-general-epoch_244.pth\n├── scripts\n│   ├── exp\n│   ├── convert_hf.py\n│   └── upload_hub.py\n├── tools\n│   ├── metrics\n├── train_data\n│   ├── example_imgs\n│   ├── motion_video\n├── inference.sh\n├── README.md\n├── requirements.txt\n```\n\n### 💻 Local Gradio Run\nNow, 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.\n```bash\n\n# Memory-saving version; More time available for Use.\n# The maximum supported length for 720P video is 20s.\npython .\u002Fapp_motion_ms.py  \npython .\u002Fapp_motion_ms.py  --model_name LHM-1B-HF\n\n\n# 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.\npython .\u002Fapp_motion.py  \npython .\u002Fapp_motion.py  --model_name LHM-1B-HF\n\n# preprocessing video sequence\npython .\u002Fapp.py\npython .\u002Fapp.py --model_name LHM-1B\n\n```\n\n### 🏃 Inference Pipeline\nNow we support upper-body image input!\n\u003Cimg src=\".\u002Fassets\u002Fhalf_input.gif\" width=\"75%\" height=\"auto\"\u002F>\n\n\n```bash\n# MODEL_NAME={LHM-500M-HF, LHM-500M, LHM-1B, LHM-1B-HF}\n# bash .\u002Finference.sh LHM-500M-HF .\u002Ftrain_data\u002Fexample_imgs\u002F .\u002Ftrain_data\u002Fmotion_video\u002Fmimo1\u002Fsmplx_params\n# bash .\u002Finference.sh LHM-500M .\u002Ftrain_data\u002Fexample_imgs\u002F .\u002Ftrain_data\u002Fmotion_video\u002Fmimo1\u002Fsmplx_params\n# bash .\u002Finference.sh LHM-1B .\u002Ftrain_data\u002Fexample_imgs\u002F .\u002Ftrain_data\u002Fmotion_video\u002Fmimo1\u002Fsmplx_params\n\n# animation\nbash inference.sh ${MODEL_NAME} ${IMAGE_PATH_OR_FOLDER}  ${MOTION_SEQ}\n\n# export mesh \nbash .\u002Finference_mesh.sh ${MODEL_NAME} \n```\n\n### Custom Video Motion Processing\n\n- Download model weights for motion processing.\n  ```bash\n  wget -P .\u002Fpretrained_models\u002Fhuman_model_files\u002Fpose_estimate https:\u002F\u002Fvirutalbuy-public.oss-cn-hangzhou.aliyuncs.com\u002Fshare\u002Faigc3d\u002Fdata\u002FLHM\u002Fyolov8x.pt\n  wget -P .\u002Fpretrained_models\u002Fhuman_model_files\u002Fpose_estimate https:\u002F\u002Fvirutalbuy-public.oss-cn-hangzhou.aliyuncs.com\u002Fshare\u002Faigc3d\u002Fdata\u002FLHM\u002Fvitpose-h-wholebody.pth\n  ```\n\n- Install extra dependencies.\n  ```bash\n  cd .\u002Fengine\u002Fpose_estimation\n  pip install mmcv==1.3.9\n  pip install -v -e third-party\u002FViTPose\n  pip install ultralytics\n  ```\n\n- Run the script.\n   ```bash\n   # python .\u002Fengine\u002Fpose_estimation\u002Fvideo2motion.py --video_path .\u002Ftrain_data\u002Fdemo.mp4 --output_path .\u002Ftrain_data\u002Fcustom_motion\n\n   python .\u002Fengine\u002Fpose_estimation\u002Fvideo2motion.py --video_path ${VIDEO_PATH} --output_path ${OUTPUT_PATH}\n\n   # for half-body video, e.g. .\u002Ftrain_data\u002Fxiaoming.mp4, we recommend to use command as below:\n  python .\u002Fengine\u002Fpose_estimation\u002Fvideo2motion.py --video_path ${VIDEO_PATH} --output_path ${OUTPUT_PATH} --fitting_steps 100 0\n   ```\n\n- Use the motion to drive the avatar.\n  ```bash\n  # if not sam2? pip install rembg.\n  # bash .\u002Finference.sh LHM-500M-HF .\u002Ftrain_data\u002Fexample_imgs\u002F .\u002Ftrain_data\u002Fcustom_motion\u002Fdemo\u002Fsmplx_params\n  # bash .\u002Finference.sh LHM-1B-HF .\u002Ftrain_data\u002Fexample_imgs\u002F .\u002Ftrain_data\u002Fcustom_motion\u002Fdemo\u002Fsmplx_params\n\n  bash inference.sh ${MODEL_NAME} ${IMAGE_PATH_OR_FOLDER}  ${OUTPUT_PATH}\u002F${VIDEO_NAME}\u002Fsmplx_params\n  ```\n\n## Compute Metric\nWe provide some simple scripts to compute the metrics.\n```bash\n# download pretrain model into .\u002Fpretrained_models\u002F\nwget https:\u002F\u002Fvirutalbuy-public.oss-cn-hangzhou.aliyuncs.com\u002Fshare\u002Faigc3d\u002Fdata\u002FLHM\u002Farcface_resnet18.pth\n# Face Similarity\npython .\u002Ftools\u002Fmetrics\u002Fcompute_facesimilarity.py -f1 ${gt_folder} -f2 ${results_folder}\n# PSNR \npython .\u002Ftools\u002Fmetrics\u002Fcompute_psnr.py -f1 ${gt_folder} -f2 ${results_folder}\n# SSIM LPIPS \npython .\u002Ftools\u002Fmetrics\u002Fcompute_ssim_lpips.py -f1 ${gt_folder} -f2 ${results_folder} \n```\n## ComfyUI Node of LHM\nWe 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\u002Fcomfyui](https:\u002F\u002Fgithub.com\u002Faigc3d\u002FLHM\u002Ftree\u002Ffeat\u002Fcomfyui) for more information!\n![](https:\u002F\u002Fvirutalbuy-public.oss-cn-hangzhou.aliyuncs.com\u002Fshare\u002Faigc3d\u002Fdata\u002FLHM\u002FComfyUI\u002FUI.png)\n\n## Contribute Needed\nWe need a comfyui windows install guide of our feat\u002Fcomfyui 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!\n\n## Acknowledgement\nThis work is built on many amazing research works and open-source projects:\n- [OpenLRM](https:\u002F\u002Fgithub.com\u002F3DTopia\u002FOpenLRM)\n- [ExAvatar](https:\u002F\u002Fgithub.com\u002Fmks0601\u002FExAvatar_RELEASE)\n- [DreamGaussian](https:\u002F\u002Fgithub.com\u002Fdreamgaussian\u002Fdreamgaussian)\n\nThanks for their excellent works and great contribution to 3D generation and 3D digital human area.\n\nWe would like to express our sincere gratitude to [站长推荐推荐](https:\u002F\u002Fspace.bilibili.com\u002F175365958?spm_id_from=333.337.0.0) and [softicelee2](https:\u002F\u002Fgithub.com\u002Fsofticelee2) for the installation tutorial video on bilibili.\n\n## More Works\nWelcome to follow our team other interesting works:\n- [LHM++](https:\u002F\u002Fgithub.com\u002Faigc3d\u002FLHM-plusplus)\n- [AniGS](https:\u002F\u002Fgithub.com\u002Faigc3d\u002FAniGS)\n- [LAM](https:\u002F\u002Fgithub.com\u002Faigc3d\u002FLAM)\n\n## ✨ Star History\n\n[![Star History](https:\u002F\u002Fapi.star-history.com\u002Fsvg?repos=aigc3d\u002FLHM)](https:\u002F\u002Fstar-history.com\u002F#aigc3d\u002FLHM&Date)\n\n## Citation \n```\n@inproceedings{qiu2025LHM,\n  title={LHM: Large Animatable Human Reconstruction Model from a Single Image in Seconds},\n  author={Lingteng Qiu and Xiaodong Gu and Peihao Li  and Qi Zuo\n     and Weichao Shen and Junfei Zhang and Kejie Qiu and Weihao Yuan\n     and Guanying Chen and Zilong Dong and Liefeng Bo \n    },\n  booktitle={arXiv preprint arXiv:2503.10625},\n  year={2025}\n}\n```\n","LHM是一个基于单张图片在几秒钟内重建可动画化人体模型的项目。它使用PyTorch实现，能够从一张静态图像生成高质量、高精度的三维人体模型，并支持后续的动作驱动。该项目的核心功能包括快速的人体重建和高效的动画生成，同时通过优化算法降低了对GPU内存的需求，使得更多用户能够在有限硬件条件下运行。LHM适用于需要快速创建数字人类角色的应用场景，如虚拟现实、游戏开发、影视特效制作等，特别适合那些希望以较低成本获得高质量3D内容的创作者。","2026-06-11 03:42:03","high_star"]