LTX 2.5 · Text to Video

The top of the LTX line, as a node

LTX 2.5 is Lightricks' newest video model, and this is its text-to-video entry: a prompt in, up to twenty seconds of video out, at anything from 720p to 4K. No checkpoint, no sampler settings, no VRAM ceiling - the model runs on PopNow's hardware and your machine just waits and downloads.

The reason to be here rather than at the LTX-2 19B node is resolution and length together. 19B gives you 1080p at best; 2.5 gives you 4K. The reason to hesitate is the bill: roughly $0.09 per second at 720p, $0.13 at 1080p, $0.19 at 2k and $0.30 at 4k. A single 20-second 4K take is about six dollars. That's a fraction of what frontier cloud video charges and roughly thirty times what the 19B node would have cost you for the same twenty seconds at 720p.

Worth saying plainly: 2.5 has far less community discussion behind it than 2.3 did. Fewer prompt recipes, fewer "here's what I learned" posts. If you want the model with a crowd around it, that's the 2.3 node. If you want the newest and the sharpest, you're reading the right page and you're on your own a bit.

What the node does

Quote, submit, poll, decode - the standard shape for this pack. The finished MP4 becomes a ComfyUI VIDEO, so it goes into Save Video or into any downstream video node. There's an idempotency key on every submission, which is why a flaky network retry doesn't quietly produce two paid renders.

The interesting bit is how little the schema asks of you. Five fields, all of them obvious:

• prompt - the picture and the action. The tooltip is short here, unlike the 19B nodes which spell out the Audio: line convention; the model page still advertises a soundtrack, so write sound cues in prose. It costs nothing and it works more often than not.
• aspect_ratio - 16:9 or 9:16. That's the whole list. No square, no 21:9 - if you need those, use the 2.3 or 19B nodes.
• resolution - 720p, 1080p, 2k, 4k.
• duration_seconds - 5 to 20, whole seconds.
• seed - fixed by default; re-queueing unchanged settings reuses the previous result instead of charging twice.

Outputs: video and task id. That task id is how you find the run in PopNow's API console later, and it's the only thing support can act on if something goes wrong.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Restart ComfyUI. Manager: search PopNow; Comfy CLI: comfy node install popnow-comfyui; or unzip a release into custom_nodes. The pack declares no dependencies beyond ComfyUI itself, so the install either works immediately or something else is wrong. Add a key from popnow.ai/api/console/keys in Settings > PopNow, or export POPNOW_API_KEY before launching - the environment wins, and it's the right call on a shared server. Restart again so the catalog populates.

Getting your money's worth

Draft everything at 720p. Twenty seconds of 720p is under two dollars and tells you whether the prompt, the camera and the length work. Only then does resolution become an interesting question.

Two settings prevent most disasters. Settings > PopNow > Max credits per run refuses any quote above your ceiling, before anything is held. And if you want variations, change one thing - the seed - and expect each queue to be a new, billed render. Seeds start on fixed specifically so an accidental double-queue doesn't cost you twice.

For errors: PopNow 40201 is an empty wallet, PopNow 40004 means that combination of settings can't be served and the message names the field, and anything ending in request_id ... should go to support verbatim. If the node itself goes missing from the menu after an update, restart - the list is rebuilt from your account's catalog at startup, so a stale menu just means a stale fetch.