Lightricks LTX 2 Pro Text-to-Video

LTX 2 Pro is Lightricks' short-form video model, and this node runs it through the WaveSpeed AI API - cloud compute, per-clip billing, zero VRAM. Where the LTX 2.3 node in this pack is the resolution-and-duration workhorse, LTX 2 Pro is the "short clip, but with synchronized audio" specialist: fixed durations of 6, 8, or 10 seconds, and audio generation built in, on by default.

It's part of the ERPK Collection, the community monorepo that wraps WaveSpeed's API (alongside OpenAI, Claude, Gemini, and Grok). No local model to download; you're renting Lightricks' model per clip.

How it works

Type a prompt (max 5000 chars - roomy), pick a duration, decide on audio, and the node sends the job to api.wavespeed.ai, polling every 10 seconds with a 15-minute timeout. The output is a video_url STRING; preview and download it with the pack's Preview Anything node. The actual rendering happens on WaveSpeed's servers.

One ERPK-family gotcha to know up front: the seed widget on this node is cache control, not reproducibility. The LTX 2 Pro endpoint has no seed parameter, so a fixed value makes ComfyUI serve the cached clip (and skip re-billing), while randomize forces a fresh generation. This is exactly why the pack ships LTX 2.3 as a separate node - 2.3's seed is real, 2 Pro's isn't, and their APIs take different parameters.

The inputs that matter

• prompt - the whole creative input. Up to 5000 chars, so you can be genuinely detailed about subject, motion, camera, and lighting.
• duration - 6, 8, or 10 seconds, as a combo. LTX 2 Pro doesn't do arbitrary lengths; pick the nearest fit. Start at 6 unless the shot needs the runway.
• generate_audio - synchronized audio, defaults to true. If you're only testing ideas, that's a per-clip surcharge for sound nobody will hear yet - flip it off while iterating.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/eRepublik-Labs/comfyui-nodes-erpk.git erpk
cd erpk
pip install -r requirements.txt

Or ComfyUI Manager → search erpk → ERPK Custom Nodes. Nothing to download; you need a WaveSpeed API key with credits, set once via right-click canvas > ERPK Settings (Settings > ERPK > API Keys). Keys are stored per-user and never land in your workflow JSON.

Common issues

• "Prompt is required" - the node validates before billing, so an empty box costs you nothing but a red error.
• The fixed-seed clip won't change - cache control again. Randomize to re-run; a fixed value deliberately reuses the last result.
• Audio you didn't want (or a bill that's bigger than expected) - generate_audio is on by default. It's a checkbox, not a conspiracy; just turn it off when you're iterating.
• 10-second clips that drift - that's the model wandering over a longer generation, not a bug. If consistency matters, 6 seconds is the safer default.

LTX's local reputation is "fast and friendly," and the cloud version keeps the friendly part. The trade is that you're on the meter. For the resolution-flexible sibling with a real seed, the LTX 2.3 text-to-video node is right next to it in the pack.