Local setup (BYO-GPU)

Run the entire pipeline on your own machine. Once installed, iteration is much faster than Colab and you can use larger/longer videos than the free tier tolerates.

Requirements

Component	Minimum	Tested on
OS	Ubuntu 22.04	Ubuntu 22.04
GPU	NVIDIA, ≥ 8 GB VRAM	RTX A2000 12 GB
CUDA	12.1	12.1
Disk	15 GB free	—
Conda	Miniconda ≥ 24	—
ffmpeg	4.x	4.4

One-shot install

git clone https://github.com/bhaveshbakshi633/VideoToSMPL.git
cd VideoToSMPL
bash scripts/install_local.sh

The installer:

1. Clones GVHMR and GMR next to this repo.
2. Creates two conda envs — gvhmr (GPU, extraction) and gmr (CPU/GPU, retargeting + GUI).
3. Installs PyTorch 2.3 + CUDA 12.1 into gvhmr.
4. Installs this repo's core package into gmr (editable).
5. Downloads ~2 GB of model weights from HuggingFace.
6. Runs healthcheck.py.

Flags:

Flag	Effect
--yes / -y	non-interactive, assume yes to prompts
--skip-weights	clone envs only, skip HuggingFace download

Directory layout after install

Projects/IL/          ← or wherever you cloned
├── VideoToSMPL/      (this repo)
├── GVHMR/
└── GMR/

Override with env vars if you placed them elsewhere:

export GVHMR_PATH=/path/to/GVHMR
export GMR_PATH=/path/to/GMR

Run

GUI (recommended for exploration)

bash scripts/run_local.sh          # http://localhost:7860
bash scripts/run_local.sh --port 8080
bash scripts/run_local.sh --share  # temporary Gradio public URL

CLI (batch / scripting)

from pathlib import Path
from core.extraction import extract_gvhmr
from core.retargeting import retarget_to_g1
from core.sanitize import sanitize_motion
from core.render import render_mujoco

pt  = extract_gvhmr("video.mp4", output_dir="out")
pkl = retarget_to_g1(pt, "out/raw.pkl")
clean = sanitize_motion(pkl, "out/clean.pkl")
mp4 = render_mujoco(clean, "out/preview.mp4")

Notebook (local Jupyter)

Any of the notebooks in notebooks/ can be run locally. Skip the cloning/install cells — those are for Colab.

Updating

cd VideoToSMPL
git pull
# re-run install if requirements changed:
bash scripts/install_local.sh --skip-weights

Uninstalling

conda env remove -n gvhmr
conda env remove -n gmr
rm -rf VideoToSMPL GVHMR GMR

Something not working?

Run the healthcheck first — it tells you exactly what's missing.

conda activate gmr
python scripts/healthcheck.py

Then consult the troubleshooting page.