LTX-2 Pro

LTX has always been the fast lane of video generation - where Wan takes minutes per clip, LTX takes seconds - and LTX-2, the 19B generation, bet the whole family on generating picture and sound together. LTX-2 Pro is the commercial tier of that bet: synchronized audio-video generation whose killer apps turned out to be lipsync and music videos rather than generic clips. This node runs it on Runware's cloud.

That "generates the audio too" part is the genuinely different thing here. Most video nodes give you silent clips you have to score separately. LTX-2 Pro can emit the clip with synchronized audio, which collapses a whole pipeline - generate, then score, then sync - into one queued job. If you've ever done the "render silent clip, then fight to line up the music" dance, this is the node that removes a step.

Inputs that matter

positivePrompt is required. Then the controls that actually shape a clip:

• providerSettings.lightricks.generateAudio - the headline toggle. On, and the model produces synchronized audio with the video. This is the LTX-2 special sauce; it's also where the cost and latency jump, so leave it off for drafts.
• audio - a UUID or URL of an audio track (2–20 seconds). Pair this with generateAudio off for music-video-style work: the model generates visuals that match your provided track instead of inventing one. This is the workflow LTX-2 actually dominates.
• duration - 6, 8, or 10 seconds (default 6). Not a free slider; the hosted spec offers these three.
• fps - 24, 25, or 50. The tooltip's reminder is accurate: 50fps is smoother but costs more processing time and money.
• frameImages - an IMAGE socket for first-frame or start/end-frame conditioning. Wire a frame and LTX animates from it.
• CFGScale - off by default, and the pack flags that this setting "has usage rules in this model." LTX likes low guidance; only enable and raise it if you're chasing stronger prompt adherence and can take the quality hit.
• width / height - 1024×1024 defaults.

numberResults (up to 4), outputFormat (MP4/WEBM/MOV), outputQuality, and the safety cluster are the standard tail.

How it works

The node sends lightricks:2@0 as a videoInference task through the Runware SDK. Video returns as a native VIDEO socket - Runware saves the clip and hands back the VIDEO type, which feeds SaveVideo or the other native video consumers in your graph. The title bar prints cost and content-check status, and it's worth watching that cost line when generateAudio is on: synchronized audio is the expensive mode.

Install and gotchas

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

Restart; API key from the Runware dashboard (Settings → "Runware API key", RUNWARE_API_KEY, or runware auth login). Paid per run, minimum top-up.

Where people stumble: they expect LTX-2 Pro to beat everything on general-purpose clips, but the KB's honest read is that LTX still loses to Wan on faces, bodies, and complex motion while holding its own on environments and camera moves - it punishes short prompts, so write full sentences about the scene. And if you feed an audio track, respect the 2–20 second window; a longer track is a guaranteed rejection, not a feature. For lipsync and audio-synced music-video work, though, this is arguably the most interesting video node in the pack.