LTX-2 19B · Text to Video

Text in, video with sound out

Twenty seconds of generated video with a soundtrack, from a prompt, on a machine that doesn't own the model. LTX-2 19B's defining bet was synchronized audio in the same network as the picture - not a foley pass bolted on afterwards - and it's the single strongest reason to reach for this node instead of a silent generator plus three more steps. The second reason is duration: twenty seconds in one shot is longer than most of what's in this price bracket.

What it's not good at: people. LTX has always been the speed tier, and the family's reputation for scenery, architecture and camera moves holding up while faces and complex body motion wobble is well earned. Treat this as the node you use for establishing shots, product reveals, mood pieces and anything where the sound is doing half the work.

How it works, and how to prompt it

The pack is a client, not an inference engine. Your prompt and widgets go into a quote call, the job is submitted with an idempotency key so a network retry never creates a second task, and the node polls until the MP4 comes back and becomes a VIDEO value you can save or feed onward.

The prompting is where beginners lose. The tooltip's hint - put the soundtrack on its own Audio: line - isn't decoration; the model responds to structure. Something like:

A slow push through a rain-slick alley at night, neon sign reflections
in puddles, the camera tilting up to a fire escape.
Audio: distant traffic, rain on metal, a low synth drone rising.

Two other habits pay off. Write long, because LTX punishes short prompts with static or chaotic results - this is a model that wants a paragraph, not a phrase. And say what the camera does, because it'll happily give you a locked-off frame if you don't ask.

The inputs

• prompt - scene, action, camera, sound on its own line.
• aspect_ratio - only 16:9 and 9:16. No square, no 21:9. Decide your format up front.
• resolution - 480p, 720p, 1080p. This is the price lever: about $0.014, $0.019 and $0.029 per second. The badge on the node updates as you change it.
• duration_seconds - 5 to 20, whole seconds.
• seed - default fixed, so the second queue of an unchanged graph reuses the first result rather than billing you again.

Outputs: video and task id. Wire video into Save Video (or a video loader later if you want the frames back as IMAGE).

Install

Manager > Custom Nodes Manager > search PopNow > Install > restart. Or:

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

There's nothing to download and nothing to compile; the pack's only real requirement is Python 3.10+ and ComfyUI's own V3 node API. Then a key: popnow.ai/api/console/keys, paste into Settings > PopNow, restart. On a server, POPNOW_API_KEY=sk-popnow-... in the environment wins over whatever's in Settings - which is also the safer option, since the key never lands in a workflow or in PNG metadata either way.

What to watch

Cost is per second of output, so a 20-second 1080p take is roughly 58 cents - cheaper than the frontier cloud models, still not free. Draft at 480p, and only re-run the winning prompt at 1080p. Env-var ceiling tip: Settings > PopNow > Max credits per run stops any run whose quote exceeds your limit, before credits are held.

Failed and refused runs are refunded; PopNow 40201 means your wallet is empty rather than your graph being broken. And if you cancel a job mid-wait, remember the job itself keeps running on PopNow and is charged if it completes - you stopped watching, not the render.

One more: the pack rebuilds its node list from your account's catalog on every start, so a missing LTX-2 node after an update usually just means "restart ComfyUI", not "reinstall".