LTX-2

LTX-2 is the model that taught people why audio matters. Lightricks' second-generation video model generates synchronized video and audio in one pass - characters talk, footsteps land, the whole clip has a sound bed that lines up - and it's the thing Wan still can't quite match. Running it locally is a 19B model with a giant text encoder and a documented VRAM headache. Running it through this node is just a prompt.

This is the API version of LTX-2, hosted on Runware's cloud through the ComfyUI-Runware pack. If you've ever wrestled the local ComfyUI-LTXVideo workflow, the difference is night and day: no Gemma text encoder to offload, no OOM fights, no distilled-LoRA juggling. You pay per render and the node prints the cost.

How it works

The node sends model AIR lightricks:ltx@2 with a videoInference task to Runware's REST API, then downloads the MP4 and wraps it as a native VIDEO output. The controls are the ones LTX users already know:

• positivePrompt and duration - both required. Duration runs 1–20 seconds, so this is a longer-form model than most video nodes.
• negativePrompt - works as expected, though it's ignored when guidance is disabled (CFG ≤ 1).
• width / height - 1024 default. LTX rewards detail; the community's number one tip is to pack the prompt with camera moves and quality keywords or you get soft, generic motion.
• steps - 40 default. CFGScale - 4. fps - 25.
• settings.enhancePrompt - on by default. LTX's automatic prompt enhancement is genuinely useful for cinematic results, but it's also the thing that rewrites your carefully worded prompt, so turn it off when you need exact control.
• numberResults - 1–4 variations.

Feed frameImages (IMAGE) for image-to-video. The output is one video socket (type VIDEO) into Preview Video or Save Video.

Installing it

One-time pack install:

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

Restart, set your key via ComfyUI Settings → Runware API key, RUNWARE_API_KEY, or runware auth login. Dependencies are just the SDK, pillow, and soundfile.

Where people get burned

Expecting Wan-tier prompt adherence. LTX's honest position in the community is "fast and great with sound, weaker than Wan on strict prompt following" - the API doesn't change that. If your renders look generic, the culprit is almost always a thin prompt or enhancePrompt mangling your intent, not the model. And note this is the base LTX-2: the fixed, much better 2.3 node is a separate model in this pack, and it's where most people land after trying 2.

The output

One video socket, type VIDEO, into a preview or save node. Older ComfyUI builds degrade it to a saved file path - update ComfyUI for the native type.