walid@portfolio:~/lab/ltx-local-video$
cd../lab
01ideaSep 2026

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.

Open weightsVideoLicensingLoRALocalModel on Hugging FaceLTX DesktopThe licenceAPI pricing
i
Renting the engine versus owning it

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.

Checked, not copied

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.

ClaimWhat the primary source says
Model and gateLightricks/LTX-2.5 on Hugging Face is real, licensed “other”, gated, and past 1.6 million downloads. The gate is why you must be signed in and have accepted terms before anything downloads.
Hardware floorExactly as described: 16GB VRAM on Windows with CUDA, or 15GB free RAM on an Apple Silicon Mac — free at launch, not installed. 160GB of disk. 16GB system RAM, 32GB recommended.
LinuxSupported for local generation with a CUDA GPU at the same 16GB floor. The guide only offers Windows and Mac, so Linux users are told to use the browser when they need not.
Text encoderGemma 4 12B, fine-tuned for LTX and bundled with the model. Google’s stock Gemma 4 is explicitly not a substitute — loading checks the encoder against the version the checkpoint was trained with.
Licence revenue lineVerbatim: entities with annual revenues of at least $10,000,000 must obtain a paid licence for any use beyond non-commercial evaluation, aggregated across subsidiaries, affiliates and companies under common control.
API costCorrect. ltx-2-5-fast is $0.09 a second at 720p and $0.13 at 1080p, so the 8-second example really is $1.04.
TrainerLinux with CUDA, 80GB+ VRAM recommended, and a low-VRAM config using INT8 quantisation for 32GB cards. train-model is a repository skill shipped in the project, not a built-in command.
Vocabulary

Five words that make the rest readable

Model — the trained thing that makes the video. The engine. LTX Desktop and LTX Studio are cars built around it.Open weights — the model files are public, so you can download and run them. The recipe and the kitchen, not a plate of food.GPU and VRAM — the graphics card and its own memory. For AI video, VRAM is the number that decides what you can run at all.Text encoder — the part that reads your prompt and works out what you meant. Here it is a 12B Gemma 4 variant, which is why long prompts survive.LoRA — a small add-on file holding one lesson: a style, a product, a character. Clips onto the engine without retraining it.
What changed

The five differences in 2.5

01·01

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.

02·02

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.

03·03

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.

04·04

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.

05·05

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.

i
Why consistency was the thing worth fixing

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.

First

Find out which path is yours — two minutes

1

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.

2

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.

3

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.

Pick one

The four routes

RouteWhat it needsWhen it is the right one
LTX Desktop16GB VRAM on Windows or Linux with CUDA, or 15GB free RAM on Apple Silicon. 160GB disk.The default. Easiest local route, with a built-in editor, and it falls back to cloud mode by itself if the machine is short.
LTX Studio (browser)Nothing. An account.Trying the model from any laptop. Note the free plan is a personal-use licence — not for client or paid work.
ComfyUI32GB VRAM and 100GB free disk.Finer control over every setting, and access to the advanced workflows. Start from the LTX-2.5 template rather than building a graph.
The APIAn account with credit. No local hardware at all.Generating videos from inside your own app, site or automation. Also the only route to ltx-2-5-pro, which is not available locally.
The local route

Installing LTX Desktop

1

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.

2

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.

3

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.

4

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.

5

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.

6

Let the first run finish

Roughly 10GB then 45GB. Hours on home broadband, once. Start it before dinner.

7

Generate the coffee-cup prompt

One subject, one light source, one camera move. If that comes back clean, the setup works.

!
Two licences, and the restriction that usually goes unmentioned

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.

The licence

What it actually says, in plain English

Under $10,000,000 in annual revenue: free to use, including commercially — run it, make client videos, fine-tune itAt or above that: a paid Commercial Use Agreement is required for any use beyond genuine non-commercial evaluation (ltxv-licensing@lightricks.com)Revenue is aggregated across subsidiaries, affiliates and companies under common control — a small brand inside a large group counts as largeYour fine-tunes and LoRAs are derivatives and stay under the same termsPublished output must be clearly disclosed as machine-generatedNo impersonating real people without consent, and no stripping watermarks, metadata or provenance dataNo building a product that directly competes with Lightricks’ ownCommercial use only: no training other models on it, beyond LTX derivativesWorldwide grant, limited only by standard sanctions law
If you use the API

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.

Model720p1080p1440p4K
ltx-2-5-fast$0.09$0.13$0.19$0.30
ltx-2-5-pro$0.12$0.17$0.25$0.39
ltx-2-3-fast$0.03$0.06$0.12$0.24
ltx-2-3-pro$0.04$0.08$0.16$0.32
!
Auto Duration decides the length, and the length is the bill

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.

The API

One request, one MP4

Swap in your key. The image-to-video endpoint takes the same shape with a public image_uri.

text-to-video.sh14 lines
curl -X POST https://api.ltx.io/v1/text-to-video \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "A wide shot of a sunlit bakery counter as a baker slides a tray of croissants onto the shelf, warm ambient chatter and the clink of trays",
    "model": "ltx-2-5-fast",
    "duration": 8,
    "resolution": "1920x1080"
  }' \
  -o video.mp4

# 8 seconds at 1080p on ltx-2-5-fast = 8 x $0.13 = $1.04
# If the file will not play, open it in a text editor — a failed
# request writes the error message where the video should be.
The API

The same call on Windows

Command Prompt needs it on one line with the inner quotes escaped.

text-to-video.cmd1 lines
curl -X POST https://api.ltx.io/v1/text-to-video -H "Authorization: Bearer YOUR_API_KEY" -H "Content-Type: application/json" -d "{\"prompt\": \"A wide shot of a sunlit bakery counter as a baker slides a tray of croissants onto the shelf, warm ambient chatter and the clink of trays\", \"model\": \"ltx-2-5-fast\", \"duration\": 8, \"resolution\": \"1920x1080\"}" -o video.mp4
Prompting

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.

Prompt 1

The setup test — single shot

Deliberately simple. If this comes back clean, your install is fine.

01-coffee-cup.txt6 lines
A close-up shot of a white ceramic coffee cup on a wooden cafe
counter in soft morning window light. Steam rises slowly from the cup
as a barista's hand slides it forward toward the camera. The camera
pushes in gently until the latte art fills the frame. Warm cafe
ambience, quiet chatter and the hiss of a milk steamer in the
background.
Prompt 2

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.

02-flower-shop.txt9 lines
A wide shot shows a small flower shop at dawn, the owner, a woman in
her fifties with short grey hair and a green apron, opening the glass
door as soft birdsong fills the street. A hard cut transitions to a
medium shot inside the shop as the same woman in the green apron
arranges red tulips into a bucket, morning light falling across the
counter; the birdsong fades and gentle acoustic guitar begins. Another
hard cut moves to a close-up of her hands tying a ribbon around a
finished bouquet; the guitar continues across the cut. She looks up and
smiles, saying warmly, "Good morning, what can I make for you?"
Prompt 3

Image to video

Keep image-to-video a single continuous shot unless you deliberately want to cut away from the photo.

03-photo-to-life.txt5 lines
Use the provided start image. The camera slowly pushes in toward the
front door of the house as late afternoon sunlight moves across the
facade. A light breeze moves the leaves of the tree in the front
garden. Quiet suburban ambience, distant birds and a car passing on a
far street.
Prompt 4

A line to camera

Exact words in quotation marks, accent named, one or two short sentences at most.

04-talking-head.txt5 lines
A medium close-up of a man in his thirties with a short beard and a
navy shirt, standing in a bright modern office with plants behind him.
He leans slightly toward the camera, raises his eyebrows and speaks in
a friendly American accent: "We answer every message within one hour.
Try us." The camera stays still. Soft office ambience in the background.
Prompt 5

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.

05-neon-sign.txt4 lines
A wide shot of a small neighbourhood bakery at night in light rain. A
bright neon sign above the window reads "OPEN" in red capital letters,
its glow reflecting in the wet pavement. A cyclist rides past slowly
from left to right. Rain patter and the low hum of the neon sign.
Prompt 6

A product reveal in two cuts

Same product, same lighting, both shots. Describe colour, material and one distinctive detail each time it appears.

06-product-reveal.txt7 lines
A top-down shot of a matte black wireless speaker sitting on a light
oak desk next to a notebook and a pen, soft daylight from a window on
the left. A hand reaches in and taps the top of the speaker, and a
small ring of white light glows around its edge. A hard cut transitions
to a low-angle close-up of the speaker as the camera slowly circles it,
the same soft daylight catching its fabric texture; a warm, deep bass
line starts playing from the speaker and continues to the end.
Making it yours

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.

Avoidable

The mistakes that cost the most time

Generating with a game or a wall of browser tabs open — they eat the same graphics memory the video needsSkipping the Hugging Face licence screen, then wondering why the model never downloadsExpecting Retake and Extend on local 2.5 — they are not there; 2.3 Fast has them, on different licence termsWriting prompts as a numbered shot list instead of one flowing paragraph that names its cutsPacking five actions into one shot — walks in, sits, opens a laptop, laughs, waves is a scene, not a shotMood words instead of detail: “cinematic, epic, stunning” says nothing, “low sun through a dusty window, slow push in” says everythingMixing light sources — pick morning window light or neon at night, not bothAssuming nothing leaves the machine: cloud text encoding sends prompt text, and anonymous usage stats are on until you turn them off
Worth knowing

The moves that save the most time

Prove it works small and short, then raise resolution one step at a time, then lengthUse the Enhance button on a thin idea and then read what it wrote — it is a free prompting lesson every timeGenerate several takes of a clip inside the timeline so you can compare without losing the one you likedRound-trip through your real editor: timelines import and export to Premiere Pro, DaVinci Resolve and Final Cut ProTreat any text in frame as a bonus; add words and logos in the editKeep a plain notes file of every prompt that produced something good, and change only the subject next time — that file is worth more than any single clip
!
Download only from the official addresses

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.