LTX-2.3 Fast

LTX's distilled tiers are the reason people say "it's not Wan, but I can do ten generations in the time Wan does one." LTX-2.3 Fast is that philosophy as a node: the 2.3 model with the step count and guidance controls stripped out, because the whole point is to render fast and iterate. On Runware's cloud that speed is the API's, not your GPU's, but the workflow benefit is the same - you burn through prompt ideas cheaply, and the node prints the per-run cost so you can watch it happen.

This is the hosted lightricks:[email protected], part of the ComfyUI-Runware pack. It keeps most of what makes 2.3 good and removes the knobs a distilled model doesn't need.

How it works

Widgets become a request with model AIR lightricks:[email protected] and a videoInference task, sent over REST through the runware-sdk; the MP4 comes back wrapped as a native VIDEO output. What changed versus the full 2.3 node:

• No steps, no CFGScale. A distilled model doesn't take them the way a base model does, so the pack doesn't pretend. You get fps (1–120, default 25) instead.
• audio is still here - the audio-to-video conditioning input (URL/UUID, merged and trimmed, max 30s) that makes LTX special.
• lora socket still present, so your Lightricks LoRAs stack the same as on the full model.
• positivePrompt, width / height (128–2048, step 64), duration (1–20s float) - required, as on 2.3.
• seed, frameImages, numberResults (1–4), settings.enhancePrompt (on by default) complete the set.

The output

One video socket, type VIDEO, into Preview Video or Save Video. Older ComfyUI falls back to a file path string; update ComfyUI for the native type.

Installing it

Same 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 stay light - SDK, pillow, soundfile.

Where people get burned

Fast is fast, not free - the per-run meter still ticks, and the temptation to iterate past the point of diminishing returns is real. Also, distilled models have their own character: fine detail can be softer than the full 2.3, so if a prompt keeps coming back mushy, that's the distillation, not a settings bug (there are no settings left to fix). For final renders worth keeping, the full 2.3 node is the safer bet; use Fast as the scratchpad it was built to be.