🅛🅣🅧 Gemma API Text Encode

Here's the single biggest pain point of running LTX-2 locally: the text encoder. LTX-2 uses Gemma 3 12B to turn your prompt into conditioning, and that thing is enormous - 22.71 GB in fp16. On launch it caused the majority of OOM errors even on high-end cards, because ComfyUI kept loading and unloading it around every generation. This node makes that problem someone else's.

What it actually does

The name is honest for once: it really does call an API. Instead of loading Gemma onto your GPU, it ships your prompt to Lightricks' hosted encoder, gets the conditioning back, and hands it to the model that's running locally. The heavy 22GB encoder never touches your VRAM. That's a huge deal on a 16GB or 24GB card - it frees up the memory that was the whole reason people were stuck. Lightricks shipped this as part of the late-January drop aimed squarely at the text-encoder complaint, and it's paired with a set of save/reuse-encoding nodes for the same reason.

The inputs that matter

• api_key - yes, you need one. It's a free key from Lightricks; this is the one setup step. No key, no encode.
• prompt - your text, multiline. LTX rewards long, specific, prose-y prompts, so write more than you think you need.
• enhance_prompt (default on) - runs Gemma's own prompt enhancement before encoding. Handy in theory, but the community verdict on the built-in enhancer is that it's basic - "just add a Show Text node to what Gemma is generating, and you'll see." A lot of people turn this off and write (or LLM-write) their own detailed prompt instead.
• ckpt_name - picks which LTX model the encoding is targeted at, so the conditioning matches.

The single output is conditioning, which wires into your sampler or guider exactly where a normal text-encode's output would.

Installing it

It's in the official Lightricks pack:

• ComfyUI Manager - Install Custom Nodes, search LTXVideo, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/Lightricks/ComfyUI-LTXVideo, then restart.

Because the encoding happens server-side, this is one of the few LTX nodes that doesn't add to your local model download - that's the point.

Common issues

Blank or errored conditioning. Almost always the api_key - check it's set and valid. This node can't fall back to a local encoder; the key is load-bearing.

Your prompt isn't landing. If enhance_prompt is on, Gemma is rewriting your text before it encodes, so what you typed isn't quite what got sent. Turn it off if you want your exact wording, and lean on a longer, more descriptive prompt - short prompts reliably underperform on LTX regardless of encoder.

You want to run fully offline. Then this isn't your node - it needs the network. Use the local Gemma encode path (and expect to feed the 22GB encoder your VRAM), or cache an encoding to disk once and reload it.