This is the step-by-step guide from the reel: the official download links, the exact hardware each route needs (PC and Mac), three ways to install it, the settings and prompt for your first clip with sound, and what it costs when you rent a GPU instead. Everything was checked on the official LTX, GitHub and Hugging Face pages on October 28, 2026.
IN ONE SENTENCELTX-2.5 is a free, open-weights video model from Lightricks that makes video and matching audio together, and it runs on your own machine if you have an NVIDIA card with about 24 GB of video memory or an Apple Silicon Mac with 32 GB or more.
It is LTX-2.5, by Lightricks. The auto captions heard "RTX 2.5". It came out in August 2026 (licence dated August 11) and it is still the newest LTX model on October 28.
Free, with a hardware bill. The model and the apps cost nothing. But a 22 billion parameter video model needs a strong graphics card. No strong card means renting one (about $0.50 to $1.19 an hour) or paying per clip on the LTX API.
"Open source" means open weights with a revenue rule. Anyone, including freelancers doing client work, can use it free while their company makes under $10 million a year. Above that you need a paid licence. It is not a classic open source licence like MIT.
Better prompt following, not perfect. The Gemma 4 text encoder does read prompts better. Each run still uses a new random seed, and LTX's own docs say on-screen text is not guaranteed. Plan on two or three tries.
Tap what you have. The chart shows how far your machine gets and which route to follow below.
Bars are to scale (80 GB is the full bar). Numbers come from LTX Desktop's own compatibility table and LTX's system requirements page.
NVIDIA · VRAM on the card
Mac · unified memory, LTX Desktop only, macOS 26+
LTX-2.5 uses the LTX-2.x Community License Agreement (dated August 11, 2026). The parts that matter for you:
Revenue counts the whole company group, not just you. The licence also has a list of banned uses (Attachment A). This is a summary, not legal advice: read the full licence before big client jobs.
| Route | Best for | Cost | Good to know |
|---|---|---|---|
| A · LTX Desktop | Beginners with a strong PC or Mac | Free | Windows, Linux, Apple Silicon. Beta app. Text encoding through the free LTX API key saves memory |
| B · ComfyUI template | Control, LoRAs, image-to-video | Free | NVIDIA only. Official minimum 32 GB VRAM, the template ships lower-VRAM files |
| C · LTX-2 repo | Coders, batch jobs | Free | About 66 GB download. Python 3.12+, CUDA 12.7+ |
| D · Runpod GPU | No strong card | $0.50 to $1.19/hr | RTX 3090 $0.50, RTX 4090 $0.89, RTX 5090 $1.19 per hour. Start with $10 of credit |
| LTX Studio | Testing prompts tonight | Free, then $15/mo | 800 one-time credits. Free plan is personal use only |
| LTX API | Pay per clip | $0.09/s at 720p | Fast model: $0.54 for 6 s at 720p, $0.78 at 1080p. Prepaid credits |
Prices checked October 28, 2026. Pakistan and India: Runpod, LTX Studio and the LTX API bill in US dollars, so use a card that allows international online payments. Big downloads: 72 GB takes about 1.5 hours on 100 Mbps and about 8 hours on 20 Mbps.
The local download takes hours on most connections. So start it, then test your prompt in the cloud while it runs:
Stuck on a step? Post a screenshot in The 1% AI Circle and we'll fix it with you.
A medium close-up shot of a young chai seller in a busy Lahore market at golden hour. He pours steaming tea from a height into a small glass, the stream arcing smoothly, then looks up at the camera and smiles. He wears a faded blue kurta and a white cap, a thin mustache, sleeves rolled up. Warm orange sunlight cuts through dust in the air, string lights blur in the background. The camera slowly pushes in toward his hands as the tea settles. We hear the clink of glass, market chatter and a distant rickshaw horn, and he says in a warm voice, "Garam chai, fresh, fresh!"
Lightricks' own free app. It downloads the model, picks settings that fit your hardware and greys out the ones that don't. Status: beta.
The LTX-2.5 files are "gated": free, but you must accept the terms first.
LTX-Desktop-Setup.exe. Mac (Apple Silicon): LTX-Desktop-arm64.dmg. Linux: the .AppImage or .deb.ComfyUI is a free node editor. LTX calls it the recommended way for most users. NVIDIA cards only for LTX.
Same as Route A step 1: accept access on the LTX-2.5 Hugging Face page. The template itself reminds you to do this first.
video_ltx2_5_t2v). There is also Image-to-Video and First-Frame / Last-Frame.The template defaults are a good first run. Paste your prompt into the prompt field and click Run. Output lands in ComfyUI/output/.
| Setting | Start with | Why |
|---|---|---|
| width x height | 1280 x 720 | Template default. Must divide by 32 |
| duration | 5 s | Short clips fit more cards |
| frame_rate | 24 | Cinematic. 25 standard, 30 smoother |
| seed | fixed | Lets you repeat a result you like |
| prompt_enhance | off | Turn on for short rough prompts (extra model download, adds 1 to 2 min) |
Best for batch jobs on a Linux box or a rented GPU. Needs an NVIDIA card (32 GB+ VRAM per LTX), Python 3.12+ and CUDA 12.7+. Commands are from the official README.
curl -LsSf https://astral.sh/uv/install.sh | shcurl -LsSf https://hf.co/cli/install.sh | bashpowershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"powershell -ExecutionPolicy ByPass -c "irm https://hf.co/cli/install.ps1 | iex"git clone https://github.com/Lightricks/LTX-2.gitcd LTX-2uv sync --extra nattenhf auth loginhf download Lightricks/LTX-2.5 diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors vae/ltx-2.5-video-vae-bf16.safetensors vae/ltx-2.5-audio-vae-bf16.safetensors latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors --local-dir models/ltx-2.5
uv run python -m ltx_pipelines.distilled \ --transformer-path models/ltx-2.5/diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors \ --text-encoder-path models/ltx-2.5/text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors \ --video-vae-path models/ltx-2.5/vae/ltx-2.5-video-vae-bf16.safetensors \ --audio-vae-path models/ltx-2.5/vae/ltx-2.5-audio-vae-bf16.safetensors \ --spatial-upsampler-path models/ltx-2.5/latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors \ --num-frames 121 --seed 42 --output-path output.mp4 \ --quantization fp8-cast --offload cpu \ --prompt "PASTE YOUR PROMPT HERE"
output.mp4 appears in the LTX-2 folder with sound. 121 frames at 24 fps is about 5 seconds.--quantization fp8-cast --offload cpu) are LTX's own memory savers; remove them on an 80 GB card for speed. On Windows, put the command on one line without the backslashes.You get a real NVIDIA card in the browser and run the exact Route B template on it. You pay only while the machine is on.
Sign up at runpod.io and add at least $10 of credit (what Runpod's own ComfyUI tutorial asks for).
Do Route B steps 3 and 4 inside that browser tab. Download your MP4s, then stop the pod in the Runpod console. A running pod keeps billing even when you are asleep.
Visual 3 · Cost per clip calculator · time your first clip, type it in
Math: hourly price x minutes / 60. The 3 minute starting value is an example, not a benchmark: time your own first clip, because speed depends on the card, the size and the length. Setup time, model download and idle time also bill, so real cost per clip is higher on day one.
Other no-GPU options. LTX Studio: free 800 one-time credits (personal use), then Lite $15 a month. LTX API: fast model $0.09 a second at 720p and $0.13 at 1080p, pro model $0.12 and $0.17, clips of 6 to 20 seconds.
This is why it needs so much memory, and why the sound always matches the picture: video and audio are made together in one pass, then sharpened.
In the ComfyUI template, a 1280 x 720 clip is drafted at 640 x 360 in stage 1. The 22B model plus the 12B text encoder is the memory cost; the int8 "lower-VRAM" files and LTX Desktop's free cloud text encoding both cut it.
From LTX's own prompting guide: one flowing paragraph, present tense, 4 to 8 sentences, under 200 words. No tag lists. Prompts written for Kling or Seedance should be rewritten into this shape.
Turn my video idea into a prompt for the LTX-2.5 video model.
My idea: [e.g. a barber shaving a customer in an old Lahore shop, for an Instagram ad]
Shot type I want: [single shot, or multi shot with 2 to 3 cuts]
Format: [9:16 vertical or 16:9]
Any spoken line: [e.g. "First cut is free." or none]
Write it the way LTX's own prompting guide asks:
1. One flowing paragraph in present tense, 4 to 8 sentences, under 200 words.
2. Cover in this order: the shot (scale and angle), the scene (light, colours, textures), the action as a sequence of verbs, the character (age, clothes, features, emotion shown through the body), the camera move, and the audio (ambience, music, dialogue in quotation marks).
3. For multi shot: name every cut ("the view cuts to..."), re-describe the person the same way after each cut, and say whether the audio continues or changes.
4. One light source logic for the whole scene. No on-screen text, logos or captions.
Then give me 2 variations: one with a different camera move, one with different lighting.
Do not add tags, bullet lists or shot numbers inside the prompts. Ask me first if my idea is missing something important.LTX-2.5's new trick from the reel: 2 to 4 connected shots in one clip, same person, same light, same voice. The yellow line is the playhead; watch what carries across each cut.
Written to LTX's guide: shot, scene, action, character, camera, audio. Swap the subject and keep the shape. Save them all as one text file with the button.
Ads and client demos. Swap the product.
A slow orbiting shot around a matte black wireless earbud case resting on a wet slate stone. The lid opens by itself and a soft white glow rises from inside as tiny water droplets slide down the stone. Cool blue rim light outlines the case against a dark studio background, with a faint reflection on the wet surface. The camera circles a quarter turn to the right and ends on a close-up of the open lid. A low electronic hum swells, then a crisp click as the lid locks open.
Explainers and faceless channels. Dialogue goes in quotes.
A medium shot of a woman in her thirties sitting at a wooden desk in a bright home office, looking straight into the camera. She has shoulder length black hair, a mustard sweater and small gold earrings. Soft window light comes from the left, a plant and a bookshelf are softly blurred behind her. She leans forward slightly, raises one hand and says in a calm, friendly voice, "Here is the one setting most people never change." The camera stays static. Quiet room tone, a faint keyboard click in the background.
Shows off LTX-2.5 multi-shot. Name every cut and say what the audio does.
A wide shot frames a rooftop in Karachi at dusk, the city glowing below, as a young man in a grey hoodie walks to the edge holding a kite. Soft wind and distant traffic fill the air. The view cuts to a medium close-up of his face, the same grey hoodie, eyes narrowed against the wind as he tugs the string; the wind continues across the cut. A hard cut jumps to a low angle shot of the orange kite climbing into the purple sky, the string snapping tight, and a gentle string melody begins over the wind.
Restaurant clients love this one.
An extreme close-up of a fork pulling apart a stack of fluffy pancakes as warm honey drips down the sides in slow motion. Steam curls up from the soft center and a pat of butter slides off the top. Warm morning light from a window on the right, shallow depth of field, a white ceramic plate on a light oak table. The camera holds still, then racks focus from the honey to the butter. Soft cafe ambience, a quiet clink of cutlery.
Describe what the camera reveals after it moves.
A smooth gimbal shot moves forward through the open front door of a modern apartment into a bright living room with floor to ceiling windows. Late afternoon sun lays long stripes of light across a light wood floor, a cream sofa and a low marble table. The camera glides past the sofa and turns slightly left to reveal a balcony with a city skyline beyond. Sheer curtains move gently in the breeze. Quiet ambient room tone and faint birdsong from outside.
Generic B-roll you can cut into talking-head reels.
A close-up tracking shot follows a person's hands typing quickly on a laptop in a dark room lit only by the screen. Lines of code scroll on the screen and reflect in a pair of glasses resting beside the keyboard. Cool blue light on the hands, deep shadows around them, a steaming mug at the edge of the frame. The camera slides slowly from left to right along the keyboard. Soft rapid typing sounds and a low synth pad.
Use with the Image-to-Video template or tool. Describe motion, not the photo.
The person in the image slowly turns their head toward the camera and breaks into a relaxed smile. Their hair moves slightly in a light breeze and the background stays still. The camera pushes in gently from a medium shot to a close-up over the length of the clip. The light stays the same as in the image. Soft outdoor ambience with distant birds.
Style words from the official list: claymation, stop-motion, 3D animation, hand-drawn.
A claymation scene of a small orange fox in a knitted scarf building a snowman in a tiny snowy village. The fox rolls a snowball across the ground, it grows bigger, and the fox pats it into place with both paws, then steps back proudly. Felt textures, soft warm window lights in the clay houses behind, gentle snowfall. The camera holds a static wide shot at the fox's eye level. A soft music box melody and the crunch of snow.
Two shots, one voice that carries across the cut.
A top down shot shows two hands opening a plain brown box on a white table and lifting out a silver smartwatch. A calm male voice says, "This is the only watch I wear now." The view cuts to a close-up of the watch on a wrist, the screen lighting up with a simple clock face, the same calm voice continuing across the cut, "Seven days on one charge." Bright soft daylight, clean white background, gentle acoustic music under the voice.
Openers and transitions.
An aerial drone shot glides slowly over a misty mountain valley at sunrise, a winding river catching the first golden light. Pine forests cover the slopes and thin clouds drift between the peaks. The camera moves forward and tilts down slightly to reveal a small wooden bridge over the river. Soft warm light, cool blue shadows in the valley. Wind, distant birds and the rush of water.
A LoRA is a small add-on file that teaches the model one style, camera move or character without retraining all of it.
Training is in the same repo as Route C (ltx-trainer). Start with the text-to-video LoRA config: it only needs short videos with captions. LTX also ships a /train-model agent you can run inside Claude Code that walks you through it. Most LoRAs trained on LTX-2.3 also run on 2.5, but test before client work, and a LoRA never works across different model families.
| You see | What it means | Fix |
|---|---|---|
| 401 / 403 on download | You have not accepted the gated model terms | Accept on the LTX-2.5 Hugging Face page, log in with a Read token |
| Out of memory (OOM) | Clip too big for your VRAM | 720p, 5 s, lower-VRAM files, close other apps, or Route D |
| Red nodes in ComfyUI | Missing custom nodes or old ComfyUI | Install Missing Custom Nodes, update ComfyUI, full restart |
| "Coming soon to your device" | LTX Desktop thinks your machine is below the floor | Mac: macOS 26 and 32 GB+. PC: more VRAM. Or Route D |
| Frame count error | Frames must be 1 + a multiple of 8 | Use 97 or 121 frames; sizes divisible by 32 |
| Different result every run | The seed is on randomize | Set the seed to fixed once you like a result |
I'm trying to run the LTX-2.5 video model locally and I'm stuck. Help me fix it one step at a time. My route: [LTX Desktop / ComfyUI template / official LTX-2 repo / Runpod cloud] My computer: [Windows 11 / Linux / Mac], GPU: [e.g. RTX 4070 12 GB], system RAM: [e.g. 32 GB], free disk: [e.g. 150 GB] What I clicked or ran: [paste the exact command or the menu steps] The exact error: [paste the full error text or describe the red node] Versions if I know them: [ComfyUI version, NVIDIA driver, Python] Rules: 1. Tell me the most likely cause first, in plain words. 2. Give me one step at a time and wait for my reply before the next step. 3. If my hardware is too small for what I tried, say so and tell me the closest setting that fits (lower resolution, shorter clip, the lower-VRAM model files, or a cloud GPU). 4. Never ask me to paste an API key or Hugging Face token.
This guide gets LTX running. Inside the community you get every AI video setup in one place, new tools every week, and real help when a download or a red node breaks on your machine. Post your first LTX clip and we'll help you turn it into content that gets clients.
Malik Muzamil · Helping creators and freelancers make money with AI · Built AI courses and programs for top 1% companies
Join our community