Installation
Prerequisites
videopython requires FFmpeg for video processing. Install it first:
# macOS
brew install ffmpeg
# Ubuntu/Debian
sudo apt-get install ffmpeg
# Windows (with Chocolatey)
choco install ffmpeg
Install videopython
Basic Installation
For basic video handling and processing:
With AI Features
To use every AI-powered feature (generation, understanding, dubbing):
[ai] is the single extra and installs every AI capability — transcription,
detection/scene/VLM understanding, source separation, translation, TTS, media
generation, dubbing, and the LLM auto-editing planner. The heavy ML dependencies
still load lazily at first use (no top-level imports under ai/), so importing
videopython stays light even with [ai] installed.
With the MCP Server
To drive editing from an MCP-capable agent, add the [mcp] extra (it needs [ai] too):
This installs the videopython-mcp console script (a stdio Model Context
Protocol server). See the
MCP Server guide.
Ollama is required for scene captioning, dubbing translation, and auto-editing
Scene captioning (SceneVLM, used by VideoAnalyzer), dubbing translation,
and the AutoEditor / MCP planner run against a local
Ollama server — there is no in-process fallback. Install
Ollama, then pull the default model:
ollama serve # start the local daemon
ollama pull qwen3.6:27b # the default vision/translation model
The model must support Ollama's structured-output format. The default
qwen3.6:27b is an Apache-2.0 vision model. Generation (image/video/speech/music), transcription, detection, and
audio classification do not need Ollama.
Dubbing TTS
The dubbing pipeline synthesizes speech with a local Chatterbox
TextToSpeech by default. To run synthesis out of process instead, inject
your own SpeechBackend into VideoDubber (e.g. a remote synthesizer).
Hardware Requirements
Image and video generation (TextToImage, TextToVideo, ImageToVideo) require an
NVIDIA CUDA GPU — these ~20–28B models raise on CPU/MPS rather than falling back. An
NVIDIA A40 or better is recommended for video generation. Music generation (TextToMusic)
runs on CUDA, Apple MPS, or CPU, and speech (TextToSpeech) on CUDA or CPU. Detection and
understanding models run on CPU (GPU optional).
Local-Only AI Runtime
videopython.ai runs locally and does not use cloud backend/API key configuration.
Practical setup notes:
- First AI run may download model weights.
- Image/video generation require a CUDA GPU; other AI models prefer a GPU (
cuda, ormpswhere supported) but run on CPU too. - Use the
deviceargument where supported to force placement.