清华GUAVA：一张照片就能搞定3D人

目录

清华GUAVA：一张照片就能搞定3D人，速度1分钟左右，流程比较长

EHM-Tracker

模型：stylematte_synth.pt

---

清华GUAVA：一张照片就能搞定3D人，速度1分钟左右，流程比较长

论文地址: https://arxiv.org/pdf/2505.03351

项目主页：https://eastbeanzhang.github.io/GUAVA/

开源代码：https://github.com/Pixel-Talk/GUAVA

视频Demo: https://www.bilibili.com/video/BV1k45AzaEpW/

依赖项：

pytorch3d

https://blog.csdn.net/jacke121/article/details/143951830

下载：

username=$(urle $username) password=$(urle $password) wget --post-data "username=$username&password=$password" 'https://download.is.tue.mpg.de/download.php?domain=smplx&sfile=SMPLX_NEUTRAL_2020.npz&resume=1' -O './data/HPS/pixie_data/SMPLX_NEUTRAL_2020.npz' --no-check-certificate --continue # PIXIE pretrained model and utilities wget --post-data "username=$username&password=$password" 'https://download.is.tue.mpg.de/download.php?domain=pixie&sfile=pixie_model.tar&resume=1' -O './data/HPS/pixie_data/pixie_model.tar' --no-check-certificate --continue wget --post-data "username=$username&password=$password" 'https://download.is.tue.mpg.de/download.php?domain=pixie&sfile=utilities.zip&resume=1' -O './data/HPS/pixie_data/utilities.zip' --no-check-certificate --continue cd data/HPS/pixie_data

EHM-Tracker

算法功能：半身动捕

python tracking_video.py -i "/tmp/gradio/99d38740e32b7465bd5274ccf3b054650b546d9061e5b9913a1387cd13dcd0ff/b002.mp4" -o "/data/lbg/project/GUAVA-main/outputs/app/tracked_driven_video" --check_hand_score 0.0 -p 0,1 -n 1 -v 0

Working directory: /data/lbg/project/GUAVA-main/EHM-Tracker

模型：stylematte_synth.pt

这个算法提供了模型包，好多模型都有：

dwpose fan hamer lmk106 lmk70 matting mediapipe pixie teaser vposer

target: src.modules.human_matting.matting.StyleMatteModel params: ckpt_fp: pretrained/matting/stylematte_synth.pt device: cuda:0

下载：

wget https://github.com/chroneus/stylematte/releases/download/weights/stylematte_synth.pth

cd /data/lbg/project/GUAVA-main/EHM-Tracker && python tracking_video.py -i "/data/lbg/project/GUAVA-main/data/nan_5s.mp4" -o "/data/lbg/project/GUAVA-main/outputs/app/tracked_driven_video" --check_hand_score 0.0 -p 0,1 -n 1 -v 0 ffmpeg -i nan1.mp4 -t 8 -c:v libx264 -c:a aac nan_5s.mp4