LTX-2 19B LoRA · Text to Video

Text to twenty seconds, with your adapter in the stack

No image needed. You write a scene, connect a LoRA, and get a clip with synchronized audio back - up to twenty seconds of it. Of everything in this pack, this is the node that behaves most like a tiny production pipeline in a single box: script, look, sound, done.

The LoRA variant costs about 25% more than the plain LTX-2 text-to-video node ($0.018/s at 480p, $0.024 at 720p, $0.036 at 1080p), and you get up to three adapters stacked. That's the deal. Whether it's worth it depends entirely on whether you have an LTX-2 19B LoRA that does something you can't prompt your way to - a specific lens character, a film stock, a recurring character. If you don't, use the non-LoRA node and keep the change.

Mechanism, briefly

Same plumbing as every node in this pack: quote, submit with an idempotency key, poll, decode. Images you connect get uploaded and swapped for popnow://f/... references; LoRA files get resolved by the PopNow LoRA node upstream - uploaded once from ComfyUI/models/loras or imported by PopNow from a Hugging Face repo, a Civitai page, or a direct link - and cached for twelve hours. The VIDEO that comes back is a plain ComfyUI value heading for Save Video.

The audio isn't a separate model pass. LTX-2 generates picture and sound in the same network, which is why the prompt structure matters: put the soundtrack on its own Audio: line, and put your LoRA trigger words at the very front, per the tooltip.

grainyfilm, a cyclist waiting at a red light in heavy rain, city bokeh behind,
camera handheld and drifting right
Audio: rain on a jacket, a wet idling engine, a distant tram bell

Long prompts, not short ones. LTX gives you static or chaos when you under-write, and no amount of LoRA scale fixes a two-line scene description.

The inputs

• prompt - trigger words first, scene and camera next, Audio: last.
• aspect_ratio - 16:9 or 9:16 only.
• resolution - 480p, 720p, 1080p. The lever on cost.
• duration_seconds - 5 to 20, whole seconds.
• seed - fixed out of the box. Same graph, same result, no second charge. Randomize when exploring.
• loras - one to three adapters, each with its own scale. Files over 2 GiB are rejected outright.

Outputs: video, plus a task id that identifies the job in the API console.

Install

Manager > PopNow > Install, restart. Or the manual route, which is two lines and no dependencies:

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

Key next: create one at popnow.ai/api/console/keys, paste into Settings > PopNow, restart. POPNOW_API_KEY in the environment takes precedence, which is what you want on any shared or headless machine. The ready-made lora template in ComfyUI's template browser shows the wiring (LoRA node → model node) if you'd rather see it than read it.

The failure modes people actually hit

Adapter does nothing. Family mismatch is the usual cause - only LTX-2 19B LoRAs work here, not LTX-2.3 files and certainly not a stills LoRA. After that, check that the trigger phrase exists in the prompt and appears early.

Import fails before generation. A gated Civitai or Hugging Face file needs CIVITAI_API_KEY or HF_TOKEN in ComfyUI's environment. Both are read from the environment only, never from a node widget, because widgets get saved into every workflow you share.

Wrong scale assumption. 1.0 is the node's default, not the adapter's sweet spot. Video LoRAs often want less; stacked adapters almost always want less each than they'd take alone.

Cost surprise. Rate × seconds × resolution adds up faster than the badge suggests, since the badge is a rate and not a total. Each run takes a quote first and holds at most that much; Settings > PopNow > Max credits per run will refuse a quote above your ceiling. And cancelling in ComfyUI doesn't stop the server-side job - it completes, and you pay for it if it succeeds.