A free AI video generator that runs offline

How to install a completely free AI video generator on your own computer and turn a text prompt or a single image into a cinematic clip, with sound effects, visual effects, camera moves and different styles built in. No subscription, no credits, no watermark, and once it is installed, no internet connection. (Free with conditions: see the licence box at the end.)

Follow one step at a time, in order. Steps 1 to 3 are a one-time setup, and the only part that needs internet. Steps 4 to 8 are how you make the videos. Steps 9 and 10 are how you run it offline and finish the edit.

The model runs on your graphics card, and the card decides how fast it runs, how big the clips can be, and on smaller cards whether a clip finishes at all. (The guide says it “decides your settings, not whether it works”.) LTX-2 is a 19-billion-parameter model, so it needs an NVIDIA card. Find yours: press Ctrl+Shift+Esc for Task Manager, open Performance, click the NVIDIA GPU and read Dedicated GPU memory. That is your VRAM. Then match it to the table. Every tier that finishes a clip produces watermark-free video; the difference is resolution, speed and, on the smallest cards, whether it finishes at all.

You also need disk space: about 45 GB for the model files, plus the app and your videos, so leave 60 GB free; Lightricks recommends 100 GB. For system memory, 32 GB is the minimum, and 64 GB helps if your card has less than 24 GB. 16 GB may work, with heavy disk use and long waits. (The guide says 40 GB of disk, 32 GB of memory recommended and 16 GB fine but slower.)

Mac with Apple Silicon: Comfy Desktop runs there, but Lightricks supports LTX in ComfyUI only on NVIDIA cards. On a Mac, Lightricks’ own LTX Desktop runs the newer LTX-2.5 locally, but only on Apple Silicon with at least 32 GB of total memory (the machine’s size, not free memory), and it needs about 72 GB of free disk; with less memory, or on an Intel Mac, local generation is unsupported and it offers only Lightricks’ paid cloud API. LTX-2.5 also comes under Lightricks’ newer LTX-2.x licence, not the LTX-2 one described at the end, and its download needs a free Hugging Face sign-in. (The guide says LTX-2 is “very slow” on a Mac. The Lab entry on LTX-2.5 linked at the top predates these figures.) Linux with an NVIDIA card follows the same steps; Comfy Desktop has a Linux version. Not in the guide.

You now have the interface. It looks like a canvas with boxes (nodes) connected by wires. You won’t need to build anything yourself: the next step loads a ready-made workflow, with most of its wiring tucked inside one big node.

When the downloads finish, the workflow is on your canvas with every box already connected. From here on you only touch a few things: the Load Image and Resize Image/Mask boxes (in the image template), the big Text to Video (LTX 2.0) node, which holds the prompt, the length, the seed and, in the text template, the size, and the Run button. (The guide says three things, and calls the big node “the prompt box”.)

The structure to keep every time: subject and action, then environment, then camera, then look and style, then “Audio:” followed by the sounds you want. Describe motion, what happens across the clip, not a still photo.

Lightricks’ own guide adds a few things. Write one flowing paragraph in the present tense, four to eight sentences and under 200 words. Name a stylised look early. Characters can speak or sing: put the words in quotation marks and name the language or accent. And clips with no speech can come out with weaker sound. Not in the guide.

The audio is generated with the picture, so you direct it like a sound designer. End every prompt with an “Audio:” line listing two to four sounds, the most important first, matched to what is on screen. (That format is the guide’s habit: Lightricks writes the sound into the paragraph and sets no number.) For visual effects, describe them as physical things happening in the scene; there are no effect settings, the words do it.

Same prompt, different look: the style words do the work. (The guide says the style words “at the end do all the work”.) Replace the look-and-style part of your prompt with one of these blocks and keep the rest the same, and you get a similar scene in a completely different style. Expect the framing and motion to shift a little too, since any change of words changes the result. (The guide promises “the same shot”.) Lightricks also finds that a stylised look works best named early in the prompt, so try moving the block to the front if the style doesn’t take.

Every finished clip is saved automatically as an MP4 with sound, in the video folder inside Comfy Desktop’s shared output folder (on Windows, under your AppData folder by default). To open it, choose ☰, Desktop Settings, Storage, and click the folder path shown under Shared Output (not the folder button beside it, which changes where your videos are saved), or press A in ComfyUI to browse your generated clips. (The guide says Documents, ComfyUI, output, and to right-click the Save Video box for Open Image Folder; that was the older app, and the menu item doesn’t exist.)

Two small print points, neither in the guide. Whenever it is online, the app checks for its own updates every 10 minutes, and that can’t be switched off; to stop it downloading and installing them on its own, untick Automatically install Desktop updates (Desktop Settings, Updates), and to stop it sending usage data, untick Share anonymous usage telemetry (Desktop Settings, Preferences). And the raw MP4 carries your prompt and the whole workflow in its metadata, so post the editor’s export, not the raw file.

The free editors have limits, not in the guide. CapCut’s free plan exports a hand-edited 1080p, 30 fps video with no watermark, as long as nothing in the project carries a Pro badge; delete the CapCut ending clip if a template adds one. DaVinci Resolve’s free version exports 1080 × 1920 H.264 on Windows and Mac (tick Use vertical resolution in Project Settings); on Windows it does so without the paid Studio’s graphics-card acceleration, so more slowly, and on Linux the free version can’t export H.264 at all.

Because the whole pipeline is local, you can make ten variations of the same Reel for different niches in an afternoon, with no fee per clip. The running cost is only electricity, which depends on how long each clip takes on your card; the real cost was the computer. (The guide says “zero cost per clip”.)