LTX Desktop: Local Video Generation With a Hard VRAM Gate
An open-source desktop app for generating videos with LTX models
At a glance
- What is it?
- Lightricks ships an Electron and TypeScript front end over LTX video models, running locally on NVIDIA GPUs or Apple Silicon and falling back to a paid API elsewhere. The interesting part is not the model list but the launch-time hardware check that decides which mode you get.
- Who is it for?
- Adopt LTX Desktop if you have a CUDA GPU with at least 16GB VRAM on Windows or Linux, or an Apple Silicon Mac with 15GB of free RAM at launch, and you want LTX 2.3 Fast specifically for its local Retake and Extend features. Do not adopt it if you need Retake or Extend on the 2.5 line, if you are on an Intel Mac, or if your machine sits near the memory floor and you cannot control what else is running when the app starts.
- Can I use it commercially?
- Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 20 days ago.
- What is it written in?
- Mainly TypeScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The problem LTX Desktop solves, and the hardware it refuses
Running a video diffusion model usually means assembling a Python environment, fetching weights, wiring a text encoder, and building some interface to drive it. LTX Desktop collapses that into an installer plus a first-run wizard. The README lists the feature set plainly: text-to-video, image-to-video, audio-to-video, video edit generation (Retake), text-to-image, and image editing, with a video editor interface and saved editing projects on top. The audience is a person who wants to cut a clip together, not maintain a CUDA toolchain.
The hardware gate is the part worth reading twice. On Windows and Linux the app requires an NVIDIA CUDA GPU with at least 16GB VRAM for local generation. On macOS it needs Apple Silicon with at least 15GB of free RAM, and it runs on MPS. Everything else, meaning Intel Macs, machines without CUDA, cards under 16GB, or systems where VRAM cannot be determined, drops to API-only mode and requires an LTX API key. That is a binary outcome decided by hardware detection, not a graceful degradation curve. If your GPU has 12GB, you are not getting a slower local render. You are getting a cloud client.
The RAM check runs once, at launch, and never again
This is the sharpest design detail in the documentation. On macOS the app checks free RAM, not total RAM, and it does so exactly once, when the Python backend process starts. It is not re-checked while the app runs. A browser with forty tabs open at launch can push a 32GB Mac into API-only mode, and closing those tabs afterward changes nothing. The README's remedy is explicit: quit and relaunch LTX Desktop.
That behaviour is defensible for a beta and frustrating in practice, because the failure is silent in the sense that the user sees API mode rather than an error explaining that memory was tight at the moment of startup. The docs do tell you how to diagnose it, which is better than nothing. But any workflow that involves opening the app alongside a memory-hungry editor will hit this. The check also means the answer to "can this machine run LTX locally" is not a property of the machine. It is a property of the machine at one instant.
Two local checkpoints with different feature sets
Local generation defaults to LTX 2.5 Fast, and LTX 2.3 Fast remains selectable in Settings > Models. Both are described as 22B distilled models. The distinction that matters is not quality but capability coverage: 2.5 Fast covers text-to-video, image-to-video, and audio-to-video, plus IC-LoRA and user LoRAs, but the README states it has no local Retake or Extend. Those two editing features are available locally only on 2.3 Fast. LTX 2.5 Pro is API-only, as is 2.3 Pro.
So the default is not the most capable local configuration. A user who installs the app, accepts the default, and then looks for Retake will not find it, and the fix is a settings change plus a different weight download. That is a reasonable trade if 2.5 Fast is genuinely faster or better on the generation tasks, which the README implies by making it the default, but it does not quantify. Anyone whose work is primarily editing rather than generation should switch to 2.3 Fast on day one and skip the 2.5 download entirely.
Text encoding is free through the API, which is an unusual split
To generate anything you must configure text encoding, and there are two paths. The first is an LTX API key, which the README says is completely free for text encoding and is recommended to speed up inference and reduce memory use. The second is a local text encoder, which is an extra download and enables fully local operation on supported hardware.
This is a deliberate hybrid. The heavy diffusion work stays on your GPU while a comparatively light encoding step goes to a server, and Lightricks does not charge for that step. The consequence is that "local generation" in LTX Desktop does not necessarily mean no network dependency. If you want the machine to work offline, you must take the local text encoder download, which the README describes only as an extra download available through the settings menu. It does not state the size or the hardware cost of that encoder, and that omission matters for anyone on a metered connection or a tight disk budget.
Install, data locations, and the disk number nobody reads
Installation is an installer from GitHub Releases, then first-run setup. Packaged installs check GitHub Releases for updates and apply them when you quit the app, so updates are not silent mid-session.
App data lands in platform-specific directories: `%LOCALAPPDATA%\LTXDesktop\` on Windows, `~/Library/Application Support/LTXDesktop/` on macOS, and `$XDG_DATA_HOME/LTXDesktop/` on Linux, defaulting to `~/.local/share/LTXDesktop/`. Model weights go into the `models/` subfolder beneath that. The Windows requirements list 160GB or more of free disk space for weights, the Python environment, and outputs. Linux and macOS entries say only "plenty of free disk space," which is less useful. Treat the Windows figure as the realistic order of magnitude across platforms.
Newer weights, including 2.5, are gated on Hugging Face. You sign in and accept the license before downloading, and the license text is fetched from Hugging Face. That means first run requires internet even if your eventual goal is offline operation.
LoRAs are a local-only feature, and the catalog says who made what
LoRA adapters steer local generation toward a style or subject and apply to text-to-video, image-to-video, and audio-to-video. They do not apply in API or cloud mode. The built-in library splits into Browse LoRAs for styles and subjects and Browse IC-LoRAs for in-context effects such as video-to-video, and each entry shows a preview, instructions, and a link to its Hugging Face page, so no manual file placement is needed.
You can also drop in your own `.safetensors` files. The compatibility rule is stated clearly: local generation uses LTX-2.5 Fast or LTX-2.3 Fast, both 22B distilled, and LoRAs labeled LTX-2, LTX-2.3, or LTX-2.5 target that family and are supported. The provenance labeling is the part worth crediting. Entries authored by LTX are marked official, and everything else carries a disclaimer that LTX does not endorse it. In a beta with a community catalog, that distinction is the difference between a user knowing what they are loading and guessing.
Alternatives and the honest boundaries of a beta
The most direct alternative is ComfyUI. It is node-graph based, so a text-to-video pipeline is something you assemble from loader, sampler, and encoder nodes and then save as a reusable graph. LTX Desktop is the opposite: a fixed feature menu (text-to-video, image-to-video, audio-to-video, Retake, text-to-image, image-to-image) behind a video editor interface with projects and a timeline. ComfyUI rewards someone who wants to rewire the sampling loop or chain unusual conditioning. LTX Desktop rewards someone who wants to cut a clip. If your task is "extend this shot by two seconds and drop in a transition," the timeline is the product. If your task is "test four schedulers against a custom conditioning stack," the timeline is an obstacle.
The limitations here are stated by the project itself. It is labeled Beta with an expectation of breaking changes. The frontend architecture is under active refactor, and the README says large UI pull requests may be declined for now, pointing at `docs/CONTRIBUTING.md`. That is an unusual thing to publish, and it is honest: the maintainers are telling contributors that the surface they would patch is moving. For a team evaluating this for production work, the relevant question is not whether the app functions but whether the interface you build habits around will survive the refactor. The project is telling you it may not.
The licence is Apache-2.0, which is permissive and permits commercial use, modification, and redistribution with attribution and a patent grant. That covers the application code. It does not cover the model weights, which are distributed separately and gated behind a Hugging Face license you must accept. Those are two different agreements, and the Apache-2.0 grant on the repository tells you nothing about what the weight license permits for your use case. Read the weight terms on Hugging Face before shipping anything commercial.
Maintenance cost is mostly disk and bandwidth. Three releases landed in the six days before this writing, which suggests active development and also means frequent re-downloads if weights change between versions. The 160GB disk figure is the number to plan around, and the update-on-quit behaviour means a relaunch can trigger a fetch you did not schedule.
Editorial conclusion
Adopt LTX Desktop if you have a CUDA GPU with at least 16GB VRAM on Windows or Linux, or an Apple Silicon Mac with 15GB of free RAM at launch, and you want LTX 2.3 Fast specifically for its local Retake and Extend features. Do not adopt it if you need Retake or Extend on the 2.5 line, if you are on an Intel Mac, or if your machine sits near the memory floor and you cannot control what else is running when the app starts. Verify before committing: launch the app with your normal workload open and confirm which mode the status shows, then check whether the gated 2.5 weights are downloadable with your Hugging Face account.
Community notes