---

name: paddle-gan description: "Route PaddleGAN tasks to focused workflows for setup, data, training, media applications, and deployment." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

PaddleGAN

Use this root only to check shared readiness and select a leaf workflow. Keep task-specific work in the selected leaf.

First checks

1. Read install and setup, choose a CPU or CUDA-enabled Paddle build, and install ppgan in the environment that will run the task.
2. Run the install checker; require GPU, ffmpeg, face, or CLIP support only when the requested workflow needs it.
3. For YAML-driven work, parse the config and overrides with the config checker before training or export.
4. If a check fails, use shared troubleshooting before entering a leaf.

Route exactly one primary task

Primary intent	Route
Train, evaluate, resume/load a checkpoint, inspect YAML, AMP, VisualDL, or distributed launch	training-configs
Run a single-image, face, portrait, latent, restoration, style, or ppgan.apps image predictor workflow	image-and-face-apps
Process video/audio, motion transfer, interpolation, video restoration/SR, First Order Motion, or Wav2Lip	video-and-audio-apps
Prepare, download, preprocess, or validate dataset folders and dataroot values	data-preparation
Export checkpoints, inspect static artifacts, or plan Paddle Inference, TensorRT, Serving, Lite, C++, or TIPC work	deployment-export

Use the workflow map for model-family routing and mixed requests.

Cross-workflow sequence

Route each stage separately: shared readiness → data preparation (if data changes) → training or app inference → deployment/export (only if a static/runtime artifact is requested). A model name does not override the stage: for example, Wav2Lip corpus preparation routes to data preparation, Wav2Lip media inference routes to video/audio, and Wav2Lip checkpoint export routes to deployment/export.

Do not run downloads, full training, heavy media inference, compilation, or deployment services merely to route a request.

Provenance and router metadata

• Repository provenance records the versioned evidence baseline and refresh checklist.
• Routing metadata is the machine-readable repo-skills-router entry for discovery.