LTX Text to Video 📝➡️🎬

The simplest node in the pack and the easiest way to get an LTX-2.3 clip without owning the hardware. Type a prompt, add an API key, run. That's the whole tutorial. No image, no audio, no model files - it POSTs your prompt to the hosted LTX API and drops an MP4 into ComfyUI's output/ folder a minute later.

Why would you bother, when the open weights are free to download? Because "free to download" is doing a lot of work. LTX-2.3 is a 22B model with a 12B text encoder; running it well locally means serious VRAM and a multi-gigabyte download. If you just want a few seconds of test footage, or you're on a laptop, renting the hosted render for a few cents a call is legitimately cheaper than your time. The flip side, same as every node in this pack: your prompt leaves the machine and lands on Lightricks' servers, and each run is metered.

What you're setting

Only two things are required: api_key (from ltx.video/api-keys) and prompt. Everything else is optional, and the defaults are sensible - but the optional list is where the interesting stuff hides:

• model - ltx-2-3-pro by default, with ltx-2-3-fast for quick tests and the legacy ltx-2-pro / ltx-2-fast if you somehow still want them.
• resolution - six options up to 4K. LTX-2.3's native portrait (1080x1920) is genuinely trained on vertical data, so don't sleep on it for Shorts/Reels.
• duration - 6–20s, but the tooltip is the real rule: pro models accept 6/8/10s, fast models take even numbers up to 20.
• fps - 24/25/48/50 for the 2.3 models, 25/50 for the older ones.
• generate_audio - defaults to on. This is LTX's signature move: the model produces synced audio along with the picture, not a silent clip you have to sound-design yourself.
• camera_motion - dolly_in, jib_up, focus_shift, etc. One dropdown that adds a lot of life to an otherwise static scene. Set to none if you want the model to just pick its own framing.
• seed - -1 for random, or fix one to reproduce a shot.

Outputs: frames (IMAGE, decoded from the MP4 for further processing), video (VIDEO - previews natively), and fps (FLOAT).

Install

Via ComfyUI Manager (search "comfyui-ltx-node"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/PauldeLavallaz/comfyui-ltx-node
pip install requests pillow

Restart and paste your key.

Where people get burned

The two things that actually matter here are the same two that matter for every LTX model: prompts need to be long and specific, and prompt adherence is the weak axis. The community consensus after months of LTX-2.3 is that seed-hunting beats prompt-polishing - generate a few short candidates on different seeds and pick the winner rather than agonizing over wording. A terse prompt like "a dog running" will come back vague; a paragraph describing the scene, the camera, the lighting, and the motion gets you a real shot. And remember that duration/fps constraints are hard - feed the API a value it doesn't accept for that model and the request just fails, taking the cost of a failed run with it.