---

name: latent-sync description: "Route LatentSync video generation, raw-video preprocessing, training, and evaluation workflows to the right sub-skill." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

LatentSync

Use this skill when the request is about the LatentSync repository: talking-head video synthesis, raw .mp4 preprocessing, U-Net or SyncNet training, or SyncNet/FVD evaluation.

Quick start

1. Run python scripts/check_env.py --check-imports --check-cuda --check-ffmpeg before a long workflow.
2. Add --check-scenedetect when preprocessing raw clips.
3. Add --check-assets when you want the bundled demo inputs or Gradio smoke path.
4. Read the narrow sub-skill before choosing configs or launch commands.

Route by task

• Generate one video, a small batch, or launch the local UI: sub-skills/inference/SKILL.md.
• Convert raw videos into aligned, AV-synced, quality-filtered clips: sub-skills/data-preparation/SKILL.md.
• Choose U-Net or SyncNet configs, build fileslists, or render training launches: sub-skills/training/SKILL.md.
• Score generated clips, validate SyncNet checkpoints, or compare real and fake sets: sub-skills/evaluation/SKILL.md.

Shared references and helpers

• references/configuration.md — repo-wide config, checkpoint, and resolution map.
• references/api-reference.md — verified shared runtime call surfaces.
• references/troubleshooting.md — cross-cutting install, import, backend, path, and checkpoint failures.
• references/repo-provenance.md — source snapshot and refresh baseline.
• references/repo-routing-metadata.json — structured router metadata for managed discovery.
• scripts/check_env.py — safe repo-root, import, CUDA, ffmpeg, scenedetect, and asset checker.

Working rules

• Always pass --repo-root to bundled helpers when the current directory is not the runtime tree.
• Keep temp directories isolated; the pipeline and metrics delete and recreate scratch paths.
• Treat CUDA as required for the core inference, preprocessing, and training workflows; CPU-only checks are useful only for limited import or FVD-style smoke paths.
• Do not route a training, preprocessing, or metric issue through inference just because the same checkpoints or demo assets appear in the stack.
• Use references/repo-provenance.md when deciding whether this skill still matches the current checkout or needs refresh.