L’interface est localisée ; les présentations et documents des Skills sont traduits automatiquement par IA dans votre langue. Le code, les commandes, chemins et termes techniques restent en version originale, et le document de l’auteur fait foi.
Explorer les Skills
Productivité · SKILL.md

draggan-ui

draggan-ui

Guides DragGAN desktop and Gradio point-based image manipulation, checkpoint selection, masks, latent-space controls, and UI troubleshooting.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentation de la Skill

GitHub

name: draggan-ui description: "Guides DragGAN desktop and Gradio point-based image manipulation, checkpoint selection, masks, latent-space controls, and UI troubleshooting." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

DragGAN UI

Use this route for desktop visualizer or Gradio tasks: launch the app, load a StyleGAN checkpoint, choose a seed, place source/target point pairs, constrain edits with masks, or debug UI/model-loading failures.

Before launch

  1. Run the root scripts/check_environment.py with the local checkout and confirm CUDA for real drag optimization.
  2. Run the root scripts/check_model_assets.py --checkpoint-dir <dir>.
  3. Use references/ui-workflows.md for the control model and launch decisions.
  4. Use references/troubleshooting.md for GUI, checkpoint, CUDA, and import errors.

The bundled launch helpers are dry-run by default:

python sub-skills/draggan-ui/scripts/launch_draggan_gui.py \
  --repo-root /path/to/DragGAN --checkpoint-dir checkpoints
python sub-skills/draggan-ui/scripts/launch_gradio_demo.py \
  --repo-root /path/to/DragGAN --cache-dir checkpoints --listen

Pass --execute only after the printed command and preflight output are correct. The desktop helper scans existing .pkl files and supports --pkl, --capture-dir, and --browse-dir. The Gradio helper requires at least one checkpoint and supports --cache-dir, --listen, and --share.

Editing model

  • A point pair consists of a source point and a target point; incomplete pairs should be reset before starting.
  • The flexible-area mask allows optimization in selected regions; the fixed-area mode protects selected regions. Reset or show the mask before diagnosing a stalled drag.
  • w optimization is faster/more global; w+ is slower but usually gives finer local control.
  • Step size/learning rate, motion lambda, feature index, r1, and r2 affect stability and locality. Change one parameter at a time.
  • Loading a new checkpoint or seed resets image, point, and mask state.

Related routes

Bundled files

Skills enregistrées