Can LTX-2.5 be run on your own computer without paying for each generation? Yes, but only on powerful hardware. This guide separates the local video model from cloud services, shows the installation of LTX Desktop on Windows and Mac, the first experiment, and the costs and limitations that short social media videos usually overlook.
As of October 23, 2026, LTX Desktop 1.3.1 (beta version). In version 1.3, the manufacturer significantly raised the hardware requirements, so older guides and videos may provide outdated thresholds. Before downloading several dozen gigabytes, check the requirements and license in the official sources linked in the text.
What is LTX-2.5
LTX-2.5 is a family of models from Lightricks that generate moving images along with synchronized sound. You can describe a scene with text, animate a photo, or overlay an audio recording. The model has open weights, meaning files that you download and run locally. However, open weights do not imply a lack of licensing. We discuss the terms below.
The easiest way to start is with LTX Desktop, the official application with a graphical interface. The application itself is open source under the Apache 2.0 license. The default local model is LTX 2.5 Fast (distilled model 22B), and there is also the older LTX 2.3 Fast available. The Pro versions operate exclusively through a paid API. More advanced users can work in ComfyUI or directly in Python, but that is a separate path with higher requirements.
LTX Desktop Requirements as of October 2026
Below are the official requirements for local generation, the same as in the documentation on GitHub and the application page.
| Computer | Recommended Requirements | What This Means in Practice |
|---|---|---|
| Windows 10/11 (64-bit) | NVIDIA card with CUDA and at least 24 GB VRAM (32 GB recommended), 32 GB RAM | Full range of resolutions and clip lengths. |
| Linux (Ubuntu 22.04+ or similar, x64 or arm64) | As above, plus NVIDIA driver installed | The application is available as AppImage and .deb package. |
| Mac with Apple Silicon (M1 or newer) and macOS 26 Tahoe | 48 GB of memory or more | Full local generation. |
| Mac with Apple Silicon and 32 GB of memory | macOS 26 | Limited mode: only 540p and 5 seconds. |
| Mac with less than 32 GB, Mac with Intel, AMD card, no NVIDIA card | Does not meet thresholds | No local generation. |
On each system, you need at least 100 GB of free space, and with additional LoRA and IC-LoRA adapters, even 160 GB or more. The application will run on macOS 14 and 15, but without local generation. Older macOS 13 will not run it.
The manufacturer also allows weaker hardware, although it does not recommend it directly. The minimum thresholds that the application checks are 9 GB VRAM on Windows with a card supporting fp8 and 15 GB on other CUDA cards (including Linux). On Mac, the threshold is 32 GB of installed memory, not free at startup. On weaker cards, the application offloads some data to RAM. According to the compatibility table, with 16 GB VRAM on Windows, about 40 GB RAM is needed (32 GB if models are on NVMe disk), and the choice of resolution and length is smaller. Below the thresholds, instead of generation tools, you will see the message Coming soon to your device.
Check Your Computer in Three Steps
Answer a few questions to find out what to expect from LTX Desktop on your hardware.
Will LTX-2.5 Work Locally on Your Computer?
What kind of computer do you have?
How much VRAM does your card have?
On Windows, you can check this in Task Manager, under Performance, GPU. Dedicated memory counts, not shared.
Do you have macOS 26 Tahoe or newer?
How much memory does your Mac have in total?
You can check this in the Apple menu, About This Mac.
Yes, full local generation
You still need 32 GB RAM and at least 100 GB of free space. We describe the installation in the Windows and Linux section.
Possible, but with limitations
On Windows, the threshold is 9 GB for cards supporting fp8 and 15 GB for others, on Linux 15 GB. The application will offer only a portion of resolutions and lengths, and with a smaller card, it needs a lot of RAM (around 40 GB with 16 GB VRAM). The manufacturer does not recommend such hardware.
First, update your system
On macOS 14 and 15, the application will run, but without local generation. If your Mac supports macOS 26, update the system and return to the memory question.
Yes, full local generation
Make sure you have at least 100 GB of free disk space. We describe the installation in the Mac section.
Yes, but in limited mode
A Mac with 32 GB of memory will only generate local clips at 540p for 5 seconds. This is sufficient for initial trials.
Local generation will not start
The application will show the message Coming soon to your device. You will have to rely on older views in Gen Space and Video Editor, which can generate through the paid API. Before clicking anything, read the section on costs.
“Free” Has Three Meanings
- No charge for each local generation. After downloading the weights, your graphics card will render the video. You still pay for hardware, electricity, and time.
- No cloud involvement. To generate video, you need to set up text encoding. A free LTX API key speeds up the process and saves memory, but then the prompt text goes to the service. If you want to work fully locally, download Local Text Encoder in Settings, Models, and select it in Settings, General, Text encoding. After configuration, the application does not need the internet.
- No licensing restrictions. This is not true. The weights are subject to the LTX-2.x Community License. Companies and individuals with an annual revenue below $10 million can use the model commercially without fees; above this threshold, a paid license is required. Sharing your fine-tuned model may also require a paid license. Details and definitions are in the full text of the license.
Paid features include video generation through the API in Gen Space and Video Editor (LTX 2.5 and 2.3, Fast and Pro) as well as Retake and Extend in Video Editor. According to the model page, LTX 2.5 Fast in the API costs $0.09 per second of video at 720p and $0.13 at 1080p, while LTX 2.5 Pro costs $0.12 at 720p and $0.17 at 1080p. Prices may change, so check the pricing before your first paid generation.
Be cautious of features that send data externally. Image generation and editing in Gen Space may go through the fal.ai service (Z Image Turbo model) if you add its key. The Enhance button with the Gemini provider and Assistant in Gemini mode sends the prompt, chat history, and thumbnails to Google. The Assistant can be switched to the local Qwen 3.5 9B model. The application also collects anonymous usage statistics, which you can disable in Settings, General, Anonymous Analytics.
Step-by-Step Installation on Windows and Linux
- Check your card in Task Manager (Performance, GPU). Note the NVIDIA model and dedicated VRAM memory. Also check how much RAM and free disk space you have.
- Go to the official LTX Desktop releases page and download the latest installer: for Windows, the file
LTX-Desktop-Setup.exe, for Linux, AppImage or .deb package. Do not download “ready packages” from random videos or drives. - Install and run the application. At the first start, it installs its own Python environment, which takes a while.
- On the Setup screen, log in to Hugging Face and accept the license for the LTX-2.5 model page. Without this, the application will not download the weights, as access to them requires agreement to the terms.
- Optionally, add a free LTX API key for text encoding, then select models to download and a folder for them. Choose a disk with ample space. The basic LTX 2.5 Fast package takes about 72 GB.
- If you did not add a key, download Local Text Encoder in Settings, Models, and set it in Settings, General. Without text encoding, you will not generate video.
Step-by-Step Installation on Mac
- In the Apple menu, About This Mac, check three things: Apple processor (M1 or newer), amount of memory, and system version. For local generation, you need macOS 26 Tahoe and at least 32 GB of memory.
- Download the file
LTX-Desktop-arm64.dmgfrom the official releases page, install, and run the application. - On the Setup screen, log in to Hugging Face, accept the license for the weights, select models, and their location on a disk with ample space.
- Set up text encoding: free LTX API key or local encoder, just like on Windows.
- If instead of tools you see the message Coming soon to your device, your Mac does not meet the thresholds (system older than macOS 26 or less than 32 GB of memory). Closing programs will not change this, as the application checks installed memory, not free memory.
The application stores settings, models, and logs in %LOCALAPPDATA%\LTXDesktop\ on Windows, in ~/Library/Application Support/LTXDesktop/ on Mac, and in ~/.local/share/LTXDesktop/ on Linux, unless you specify a different folder for models during configuration. Before deleting anything there, check if there are materials you want to keep.
The application checks for updates automatically at startup and every 4 hours. On Mac, it installs them after closing the program; on Windows and Linux, it asks for permission and requests a restart.
First Video: Simple Exercise
Select Text to Video and the LTX 2.5 Fast model. The default settings are 540p, 8 seconds, and 24 frames per second, and it's best to start from them. For starters, one character, one action, and a stationary camera. Increase the resolution and length only when the motion looks good. A long prompt with many twists heavily burdens the computer, and the result is hard to evaluate later.
One continuous shot: a small paper crane stands on a wooden desk by the window on a rainy afternoon. A gust from the open window gently moves the crane's wings; then the paper figure shifts a few centimeters across the surface. The camera is stationary, close-up, soft light from the left, raindrops on the window in the background. Calm pace. Only light rain and the rustle of paper can be heard. No people, titles, music, or cuts.
Watch the result from three angles. Does the crane have the same number of wings throughout? Is the movement smooth and gradual? Does the sound match the scene? If the model loses the shape of the paper, limit the action to one small movement.
Animate Your Own Photo
In Image to Video, you add your own photo or illustration as the first frame. You can also add an ending frame, and the model will fill in the motion between them. This helps maintain composition. Describe the motion of what is already visible instead of adding new objects. For a photo of a wet street, simply say: “the camera is stationary, one tram passes by, reflections ripple in the puddles.” Compare the first frame of the result with the photo. If the buildings have already changed shape at the start, simplify the prompt.
What LTX 2.5 Fast Can Do Locally
| Function in LTX Desktop | Status |
|---|---|
| Text to Video with sound | Locally on supported hardware. |
| Image to Video, Audio to Video | Locally. |
| Retake (regenerate a segment of 2–10 s) and Extend (add 4–12 s) | Locally in the main view on stronger hardware; not in the mode that offloads data to RAM on weaker cards with Windows. In Video Editor, these same functions go through the paid API. |
| LoRA with styles and camera movements, IC-LoRA (e.g., day to night, denoising, colorization) | Only locally. The adapter must match the LTX-2, 2.3, or 2.5 model. |
| LTX 2.5 Pro | Only API, paid generation in the cloud. |
| Static images in Gen Space | Locally after downloading the image model or through fal.ai; check which mode is enabled. |
If You Have a Powerful Computer and Want More Control
The official LTX-2 repository provides ready pipelines in Python, and ComfyUI-LTXVideo nodes allow you to work in ComfyUI. This option is for those who want to set models, adapters, length, input frames, and rendering quality themselves. The files for a quick start in Python take about 66 GiB in total. For ComfyUI, the manufacturer recommends a CUDA card with at least 32 GB VRAM and 100 GB of free space. With less memory, the card options --quantization fp8-cast and --offload cpu help, at the cost of speed. If you just want to make your first short clip, do not start down this path.
Common Issues
| Symptom | What to Check |
|---|---|
| Message Coming soon to your device | NVIDIA card VRAM, macOS version (26 needed), or Mac memory (minimum 32 GB). A Mac with Intel does not have local mode. |
| Error downloading weights | Login to Hugging Face, accepted license, disk space, and stable connection. |
| No higher resolutions or longer clips | The application hides options that your hardware cannot handle. This is not an error. |
| Generation very slow | Shorter clip, lower resolution, slow RAM, Fast model instead of Pro. |
| Unexpected charge | Did the generation, Retake, or image go through the API? Check the mode and pricing before clicking. |
| Prompt and files were supposed to be private | Enable local text encoder, do not use Gemini or fal.ai, switch Assistant to local model. |
Is It Worth It?
Yes, if you have supported hardware and want to experiment with AI video without paying for each attempt. The biggest advantage is working on your own weights and full control over what comes out of the computer. The biggest costs are memory, disk, and rendering time. On a laptop without a strong NVIDIA card or on a Mac with 16 GB of memory, the phrase “free on every computer” is simply misleading.
For further work, our guide on framing and focal lengths and character consistency card will be useful. Both will help evaluate the result regardless of whether you generate locally or in the cloud.
Frequently Asked Questions
Does LTX Desktop work on a Mac with 16 GB of memory?
No, not in local mode. Since version 1.3, the threshold is 32 GB of installed memory and macOS 26. The application will run on a weaker Mac, but you can only generate through the paid API.
Do I need an API key to get started?
No. The LTX API key is not required. However, you need to set up text encoding: either with the free LTX key (the prompt then goes to the cloud) or with the local encoder downloaded in the settings.
Can I use videos from LTX-2.5 commercially?
The LTX-2.x license allows free commercial use if your company's or your annual revenue is below $10 million. Above this threshold, a paid license is required. Read the full text of the license before using it in business.
Is an AMD card sufficient?
No. LTX Desktop generates locally only on NVIDIA cards with CUDA (Windows and Linux) and on Macs with Apple processors.
How much disk space will LTX Desktop take up?
The manufacturer recommends at least 100 GB of free space, and with additional LoRA and IC-LoRA adapters, 160 GB or more. The basic LTX 2.5 Fast package is about 72 GB.
Sources
| Source | Type | Status / Reliability |
|---|---|---|
| LTX Desktop on GitHub (README, requirements, models) | Official manufacturer repository | Checked October 23, 2026, version 1.3.1 |
| LTX Desktop: Setup and settings | Official documentation | Checked October 23, 2026 |
| LTX Desktop: hardware compatibility table | Official documentation | Checked October 23, 2026 |
| LTX Desktop: API keys, costs, and privacy | Official documentation | Checked October 23, 2026 |
| LTX Desktop page | Official manufacturer page | Checked October 23, 2026 |
| LTX-2.5 model page (license, API prices) | Official manufacturer page | Checked October 23, 2026; prices may change |
| LTX-2.5 model card on Hugging Face | Official weights repository | Checked October 23, 2026 |
| LTX-2.x Community License | Model license | Binding text |
| LTX-2 repository (Python) | Official manufacturer repository | Checked October 23, 2026 |
| ComfyUI-LTXVideo | Official manufacturer repository | Checked October 23, 2026 |
As of October 23, 2026. This guide is for informational purposes only and does not constitute legal or licensing advice. Hardware requirements, prices, and licensing terms are set by the manufacturer and may change, especially since LTX Desktop is in beta.
Comments (0)
No comments yet. Be the first!