Lightricks/LTX-Desktop

LTX Desktop – What it is

LTX Desktop is an open‑source desktop application that lets you generate and edit videos (and images) with LTX‑branded generative models. It runs locally on powerful Windows/Linux NVIDIA GPUs or Apple‑silicon Macs, and falls back to a paid cloud API when the hardware isn’t sufficient.

Core capabilities

Capability	How it works
Text‑to‑video	Feed a prompt; the local LTX 2.5 Fast (default) or LTX 2.3 Fast model creates a video.
Image‑to‑video / Audio‑to‑video	Same models accept an image or an audio clip as additional conditioning.
Video edit (Retake / Extend)	Only the older LTX 2.3 Fast model supports local “Retake” edits; the newer LTX 2.5 Pro version does it via the cloud API.
Text‑to‑image & image‑to‑image	Integrated image generation/editing using the same backend.
LoRA support	Load LoRA adapters (style/subject tweaks) from a built‑in library or your own `.safetensors` files – works for local video generation only.
Prompt enhancement	Catalog‑aware rewriting via a local module or the Gemini API to improve prompts.
Video editor UI	Timeline‑based editor for assembling, trimming, and filling gaps in generated clips.

Local vs. API mode

Platform	When you get local generation	When you get API‑only
Windows + CUDA GPU (≥16 GB VRAM)	Downloads model weights and runs on the GPU.	If GPU is missing, VRAM < 16 GB, or driver unknown – requires an LTX API key.
Linux + CUDA GPU (≥16 GB VRAM)	Same as Windows.	Same fallback.
macOS Apple Silicon (≥15 GB free RAM)	Runs on Apple’s MPS backend.	Low free RAM or Intel Macs → API‑only.

In API‑only mode you can still generate videos, but you’re limited to the resolutions/durations the cloud service offers and you must pay for the Fast or Pro tiers.

Getting started

Download the installer from the repository’s Releases page.
Run the installer – it creates an Electron‑based UI and a Python FastAPI backend.
On first launch you may need to sign in to Hugging Face to accept the license for the gated 2.5 model weights.
Configure a free LTX API key (for cloud text encoding and optional cloud generation) in Settings → API.
Choose Local or API mode in the UI; the app will automatically detect hardware and free RAM.

All app data (settings, downloaded models, logs) lives in the standard per‑OS data folder:

Windows: %LOCALAPPDATA%\LTXDesktop\
macOS: ~/Library/Application Support/LTXDesktop/
Linux: ~/.local/share/LTXDesktop/

Development & contribution

The repo is split into three layers:

Renderer – React + TypeScript UI (frontend/).
Electron – native window, file dialogs, and process management (electron/).
Backend – Python FastAPI server that talks to the GPU and external APIs (backend/).

Quick‑start for contributors:

# install Node, uv, Python 3.13+, Git
pnpm setup:dev   # installs deps, creates virtualenv, builds UI
pnpm dev         # launches Electron + backend in dev mode

Additional scripts exist for debugging, type‑checking, running backend tests, and performance validation.

License & privacy

Code – Apache‑2.0.
Model weights – downloaded separately from Hugging Face; they carry their own licenses.
Telemetry – optional anonymous usage stats (app version, platform, random ID). Can be disabled in Settings → Anonymous Analytics.

Who might find this useful?

Creators who need a desktop‑only workflow for quick video prototyping without sending data to the cloud.
Researchers experimenting with LoRA adapters on video diffusion models.
Teams that want a single UI for text‑to‑video, image‑to‑video, and basic video editing, with the option to fall back to a paid API when hardware is insufficient.

Bottom line: LTX Desktop bundles a modern video‑diffusion pipeline into a user‑friendly Electron app, offering both fully local generation on high‑end GPUs/Apple Silicon and a cloud fallback for less capable machines. It’s a genuine, actively‑maintained AI‑focused project.