The engine on your own machine
Running an open-weights video model locally — what the hardware actually has to be, and what the licence lets a business do with the output
LTX-2.5 turns a written description into a clip with sound, and Lightricks published the weights — so it runs on hardware you own rather than being rented a clip at a time. Multi-shot with a consistent character is the headline. The parts that decide whether it is usable are less glamorous: a hard VRAM floor, 160GB of disk before the first frame, and a licence with a revenue line and a restriction most summaries leave out.
On a video website you rent the engine by the clip: someone else sets the price, and can change it or switch it off. Open weights means the engine sits on your machine — nobody meters it, and with local settings the footage never leaves the building. The trade is that the machine now has to be strong enough, and that is a specific number rather than a vibe. Everything below is about which number applies to you.
Verified against the repositories and the licence on 23 September 2026
Requirements read from the LTX Desktop README, prices from the published API pricing page, licence terms from the licence text itself.
Five words that make the rest readable
The five differences in 2.5
Multi-shot in one generation
One prompt can produce several connected shots with cuts between them — wide, then close-up, then reaction — holding the character, setting, lighting, voice and style across the cuts. Getting the same person to look the same in shot two used to be the hardest part of this.
Prompt following
The Gemma 4 encoder keeps more of a complicated prompt: several people, what each is doing, the lighting, the camera move. The claim is fewer retries, not zero.
A cleaner final pass
A diffusion video decoder gives sharper faces in close-up, more readable signage and fewer smears on fast motion. It costs decode time and VRAM.
Sound in the same pass
Audio is generated alongside the picture, so footsteps, ambience and spoken lines arrive already in sync rather than needing to be laid over afterwards.
Auto Duration
The model reads the action described and picks a clip length, up to 20 seconds. Convenient locally; on the API it means you are billed for a length you did not choose — see the pricing note below.
The multi-shot feature reads like a technical nicety and is not. A single beautiful five-second clip is a screensaver. Three connected shots of the same person, in the same room, with the same voice, is a scene — and a scene is the smallest unit you can actually post, pitch or sell. That is the whole difference between a demo and a deliverable.
Find out which path is yours — two minutes
On Windows
Ctrl + Shift + Esc opens Task Manager. Performance tab, then GPU in the left column — if there is more than one, pick the NVIDIA. Read Dedicated GPU memory. That number is your VRAM, and it decides everything that follows.
On a Mac
Apple menu, About This Mac. M1 or newer means Apple Silicon and the local route is open; Intel means it is not. Then check free memory, not total: Activity Monitor, Memory tab. The app wants 15GB free at the moment it launches, and your browser is usually what is eating it.
Then check disk
160GB free before you start. The first run pulls roughly 10GB of software and 45GB of model files, and running out halfway is the most irritating way to lose an evening.
The four routes
Installing LTX Desktop
Update the NVIDIA driver first (Windows and Linux)
An out-of-date driver is a common reason a local AI app refuses to start. Mac users skip this.
Create a free Hugging Face account
The model files are gated. You have to be signed in and have accepted the licence before the download can begin — this is the step people skip and then wonder why nothing happens.
Download from the official releases page only
github.com/Lightricks/LTX-Desktop/releases/latest. Windows takes the .exe, Mac the arm64 .dmg. If Windows SmartScreen objects, More info then Run anyway — but only for a file you downloaded from that page yourself.
Sign in and accept the licence in the app
This is what releases the model files. The licence is the LTX-2.x Community License, covered below.
Create a free LTX key and paste it in
The key handles prompt encoding in Lightricks’ cloud, which is free, faster, and frees graphics memory for the video itself. The trade is that your prompt text leaves your machine while the video does not. If prompts are confidential, switch on the local text encoder instead — about 26GB more to download.
Let the first run finish
Roughly 10GB then 45GB. Hours on home broadband, once. Start it before dinner.
Generate the coffee-cup prompt
One subject, one light source, one camera move. If that comes back clean, the setup works.
The app and the model are licensed separately: LTX Desktop is Apache 2.0, the model files are under the LTX-2.x Community License. There is also a second community licence — the older LTX-2 agreement from January 2026 covers LTX 2.3, which matters because the usual advice for Retake and Extend is to switch to LTX 2.3 Fast, quietly moving you onto different terms. And a restriction the plain-English summaries drop: for commercial use you may not use the model to train, improve or fine-tune any other model or AI system, other than derivatives of LTX itself. Anyone planning to generate a synthetic training set with it should read that clause before starting, not after.
What it actually says, in plain English
The full price list, per second of video
From the published pricing page. The guide quotes the two fast-model figures; the rest of the grid is worth seeing, because 2.3 is a third of the price and the resolution step is steep.
Because API charges are per second of generated video, letting the model choose the duration means letting it choose the invoice. The pricing page is explicit that with automatic duration you are charged for whatever length it produces, and that prepaid accounts have credits held against the longest duration your resolution and frame rate allow until the job finishes, with the remainder released afterwards. Pin the duration when you care about the cost. Worth knowing too that there is an async v2 endpoint alongside the v1 sync one used below, and an audio-to-video mode the guide does not cover.
One request, one MP4
Swap in your key. The image-to-video endpoint takes the same shape with a public image_uri.
The same call on Windows
Command Prompt needs it on one line with the inner quotes escaped.
The recipe the prompts all follow
Shot type, then the scene and the light, then the action in order, then who is in it, then how the camera moves, then the sound. Present tense, one flowing paragraph. Not a numbered shot list — naming cuts in words works better than “Shot 1, Shot 2”.
For multi-shot: keep it to two to four shots, name each cut explicitly, re-describe the person every single time in the same words, and say whether the music carries across the cut or changes. “The woman in the green apron” in every shot is what holds the character together; a fresh description invites a fresh person.
The setup test — single shot
Deliberately simple. If this comes back clean, your install is fine.
Multi-shot — the headline feature
Three shots, one character, music carrying across the cuts, a spoken line at the end. Every trick in one prompt.
Image to video
Keep image-to-video a single continuous shot unless you deliberately want to cut away from the photo.
A line to camera
Exact words in quotation marks, accent named, one or two short sentences at most.
Testing the sign rendering
One short word, big and prominent, is the best case. Longer words, brand names and prices still belong in your editor.
A product reveal in two cuts
Same product, same lighting, both shots. Describe colour, material and one distinctive detail each time it appears.
Two levels of fine-tuning
The first level needs no training at all: LTX Desktop has a built-in LoRA browser, and Lightricks publishes a substantial official set alongside the community ones — spatial upscaling, clean plate, deblur, day-to-night, slow-motion control, cinemagraph, pose control, in- and out-painting, HDR. Someone may already have made the look you were about to train for, which is why browsing comes before training. LoRAs only work in local mode, and must be built for LTX-2, 2.3 or 2.5 — anything else simply has no effect.
The second level is training your own on your product, mascot or house look, using the trainer shipped in the LTX-2 project. That wants Linux with CUDA and 80GB of VRAM for the standard config, with a low-VRAM configuration using INT8 quantisation for 32GB cards like a 5090 — or a rented GPU box by the hour. The material is a folder of your own clips with written captions, and the trainer includes tools that split long videos into scenes and caption them for you. The project also ships a train-model skill you can run inside Claude Code, which walks the steps and asks before anything expensive starts.
The mistakes that cost the most time
The moves that save the most time
The official homes are ltx.io, console.ltx.io, docs.ltx.io, huggingface.co/Lightricks and github.com/Lightricks. Search results for a model this popular fill up with lookalike domains carrying the same name, and an installer or a set of model weights is exactly the kind of file you do not want from a stranger. Check the address bar before you click Run anyway on any security warning.