LTX 2.5 Text To Video

Text to video is the simplest node in the LTX 2.5 family and the easiest to underestimate. No image to load, no audio to prep - just a prompt and a model dropdown, and a finished clip comes back. For anyone who's ever watched their local GPU choke on a video model, this is the relief valve: the generation happens on Lightricks' servers, not your machine.

It's the hosted tier of LTX 2.5, a partner node that bills per second through your Comfy account and landed in ComfyUI core in August 2026. Where the image and audio-to-video variants anchor on an input, this one is pure creation - which makes it the natural starting point when you just want to see what an idea looks like moving.

How it works

The prompt goes to the LTX 2.5 API, the model generates the clip, and the node downloads it. The model dropdown is the actual control panel: choose LTX-2.5 (Fast) or LTX-2.5 (Pro), then inside it set the duration (2 to 20 seconds), resolution (up to 3840x2160 on Fast, 1080p on Pro), fps, and whether to generate a matching AI audio track (generate_audio, on by default). Fast is the budget draft tier; Pro is higher fidelity but capped at 10 seconds and 1080p.

The per-second rate differs between tiers and resolutions, and the node shows a live price badge so you can see what a "20 second, 4K" request is going to cost before you click go. That transparency is the best feature - there's no mystery bill after a long render.

The inputs that matter

• model - Fast or Pro, with duration/resolution/fps/audio nested inside.
• prompt - the whole brief, up to 10,000 characters. Say what's in frame, how it moves, and if you want sound, what it should be.
• seed - the "force a re-run" control; results are nondeterministic regardless of what you set.

Output is a single VIDEO.

Gotchas

The same duration rule as the image node applies: over 10 seconds means 720p/1080p at 24/25 fps - the node errors if you've asked for 4K at 50 fps for 20 seconds, and Pro can't go past 10 seconds at all. LTX's local reputation is "fast but a tier below Wan on faces and complex motion," and the hosted 2.5 tier closes a lot of that gap - but it still punishes thin prompts. Give it a real scene description, or you'll get a pretty, generic pan. And since it's cloud-billed, iterate with short cheap clips before you commit to the long expensive one.