NV-Generate · Synthetic Medical Imaging

A unified Hugging Face Spaces demo for NVIDIA Medtech's three open-weight 3D medical image generators, all built on the MAISI-v2 rectified-flow architecture (~30 inference steps each).

Features

• Hero card landing → per-model workspace.
• niivue WebGL multiplanar viewer (axial / coronal / sagittal + 3D render).
• Paired 132-class anatomy mask overlay for CT (with deterministic per-label colormap + named legend).
• Window/Level presets (Soft Tissue / Lung / Bone / Brain) on the CT viewer.
• Direct NIfTI download of every generated volume + mask.
• ZeroGPU support for HF Spaces (@spaces.GPU decorator).

Local development

pip install -r requirements.txt
pip install -r repos/NV-Generate-CTMR/requirements.txt
python app.py                           # http://localhost:7860

app.py auto-clones the upstream NV-Generate-CTMR inference repo into ./repos/ on first run (no separate pre-build.sh needed). Weights are downloaded lazily from the Hugging Face Hub the first time each model is exercised, then cached.

Hugging Face Spaces deployment

Recommended hardware: ZeroGPU (A10G or H100), since each model needs ~16–80 GB VRAM depending on volume size.

1. Create a new Space on huggingface.co with sdk: gradio (already set in this README's frontmatter).
2. Push this repository:

   git remote add space https://huggingface.co/spaces/<your-username>/nv-generate
   git push space main
   

3. In the Space's Settings → Hardware, select ZeroGPU.
4. First build will install dependencies + clone the upstream repo. First generation on each model downloads weights into a persistent cache.

License

Not for clinical diagnostics. This is a research demo for synthetic data generation only.

Credits

Built on the MAISI framework:

• MAISI-v1 — WACV 2025 paper
• MAISI-v2 — AAAI 2026 paper

NVIDIA Medtech, University of Zurich, Istanbul Medipol, Forithmus.