screencastgen

Models

Model download management for ML dependencies.

Source: screencastgen/models.py

Overview

Manages downloading and caching of ML model checkpoints. Supports pre-downloading models so that pipeline runs don’t need network access.

Dataclass

`ModelPackageSpec`

Metadata for a downloadable model package.

Functions

`register_model_download_args(parser)`

Registers repeatable --backend and --package selectors, --all, and target-specific download arguments. Current packages are whisperx and latentsync; Qwen is selected as a backend.

`download_selected_models(args) -> int`

Main orchestrator. Downloads models based on CLI flags.

`_get_cache_dir() -> str`

Returns the model cache directory.

`_download_whisperx(args)`

Downloads WhisperX model checkpoints.

`_download_latentsync(args)`

Downloads LatentSync checkpoints via LatentSync Provider.

Dependencies

Models
├── TTS Registry            (get_backend_spec, download registration)
├── TTS Base                (BackendArg)
├── LatentSync Provider     (download_latentsync_checkpoints)
└──▶ consumed by CLI (download-models subcommand)

Usage

screencastgen download-models --backend qwen --model 1.7B
screencastgen download-models --package whisperx --package latentsync
screencastgen download-models --all

Qwen downloads use Hugging Face snapshot caching without loading the TTS model into memory. A local path supplied through --model is accepted without downloading.

screencastgen

Models

Overview

Dataclass

ModelPackageSpec

Functions

register_model_download_args(parser)

download_selected_models(args) -> int

_get_cache_dir() -> str

_download_whisperx(args)

_download_latentsync(args)