LGG100/linglongimage

linglong-live Docker image

Docker image of the monocular RGB → LingLong 2.0 real-time motion retargeting system (OpenLoong 大师赛第一期 · 赛题一 · 复赛), built for RTX 50xx (Blackwell, sm_120): CUDA 12.8 base, torch 2.7.1+cu128. Model weights are bundled.

Usage

Host: Linux, NVIDIA GPU (≥ 8 GB), nvidia-container-toolkit.

hf download LGG100/linglongimage linglong-live.tar --local-dir .
docker load -i linglong-live.tar          # -> linglong-live:latest

# webcam + window
xhost +local:root
docker run --rm -it --gpus all --ipc=host --network host --device /dev/video0 \
  -e DISPLAY=$DISPLAY -v /tmp/.X11-unix:/tmp/.X11-unix \
  -v linglong_cache:/cache -v $PWD/out:/out linglong-live --source 0

# headless, video file
docker run --rm --gpus all --ipc=host -v linglong_cache:/cache -v $PWD/out:/out \
  -v $PWD/input.mp4:/in.mp4:ro linglong-live --source /in.mp4 --headless \
  --record /out/out.mp4 --log /out/log.csv

The first frame compiles GPU kernels (cached in the linglong_cache volume).

Bundled third-party materials

• SAM 3D Body weights (facebook/sam-3d-body-dinov3), redistributed under the SAM License — see LICENSE_SAM.
• YOLO11m-pose (Ultralytics), AGPL-3.0.
• MoGe-2 ViT-S (Ruicheng/moge-2-vits-normal).
• LingLong 2.0 URDF and meshes (official competition release, unmodified).