screencastgen

LatentSync Provider

High-quality lip-sync video generation using LatentSync.

Source: screencastgen/providers/lipsync/latentsync_provider.py

Overview

Generates lip-synced face videos by running LatentSync in a subprocess. Supports quality presets and manages session caching for performance.

Presets

`LatentSyncPreset` (Dataclass)

Preset	Resolution	Description
`small`	256px	Fast, suitable for small face overlays
`quality`	512px	High quality, larger face overlays

Each preset specifies:

config_candidates — Config file paths to search
checkpoint_candidates — Model checkpoint paths
guidance_scale — Diffusion guidance scale
inference_steps — Number of inference steps

Function

`run_latentsync(video_path, audio_path, output_path, device="auto", preset="quality")`

Finds the LatentSync installation directory (via env var or path search)
Selects config and checkpoint for the preset
Spawns a worker subprocess (latentsync_worker.py)
Manages session caching and timeouts

`download_latentsync_checkpoints(args)`

Downloads LatentSync model checkpoints. Called by Models.

LatentSync Discovery

Finds the LatentSync root directory by:

Checking LATENTSYNC_ROOT environment variable
Searching common paths

Dependencies

LatentSync Provider
├── latentsync         (external installation)
├── subprocess         (worker process)
├── TTS Base        (resolve_device)
└──▶ registered in Lipsync Registry
     └──▶ called by Lipsync Facade