---

name: opengvlab-draggan description: "Routes DragGAN users to the browser demo and Python API workflows for point-based image editing, checkpoint loading, and CUDA troubleshooting." disable-model-invocation: true metadata: disco-role: operating license: NO_LICENSE

DragGAN

Use this skill when a task mentions DragGAN, the Gradio demo, point dragging, pretrained checkpoint selection, or scripted latent/image generation.

Quick start

1. Install a CUDA-enabled PyTorch/TorchVision pair that matches your platform.
2. Install this package: python -m pip install "draggan==1.1.6".
3. If you already have a checkout open for editing, python -m pip install -e . is also fine.
4. Run scripts/check_install.py --mode web before launching the browser UI, or scripts/check_install.py --mode api before scripting against the package.

Routes

• sub-skills/web-demo/ for the browser UI, checkpoint switching, point dragging, and saved image/video outputs.
• sub-skills/python-api/ for draggan.draggan and draggan.utils usage from your own code.

Shared references

• references/deployment.md for install and launch patterns, including the bundled launcher and Docker notes.
• references/checkpoints.md for checkpoint names, cache location, and auto-download behavior.
• references/troubleshooting.md for cross-cutting import, backend, and checkpoint failures.
• references/repo-provenance.md when checking whether this skill still matches the current repository snapshot.
• references/repo-routing-metadata.json for the route metadata used by repo-skills-router.

Important limits

• The verified drag loop is CUDA-only in this snapshot.
• The mask tab shown in the UI is not enforced by the current optimizer.
• Uploaded-image inversion is experimental and may fail; prefer the seeded-image workflow unless you are specifically debugging that path.
• If the task is only about environment problems, read the root troubleshooting reference before choosing a route.

Route selection hints

Choose web-demo when the request talks about the browser demo, clicking handle and target points, choosing checkpoints, saving outputs, or exposing the app on a host/port.

Choose python-api when the request talks about load_model, generate_W, generate_image, drag_gan, latent editing, or writing a script around the package.

If the task is unclear, inspect references/deployment.md first, then route to the sub-skill that matches the user-facing workflow.