🅛🅣🅧 LTXV Save Conditioning

The LTX-2 text encoder is Gemma 3 12B - 22.71 GB in fp16, the single biggest source of OOM errors and load-time misery on consumer cards. This node is half of the cure. Encode your prompt once, save the result to disk, and you never have to run that 22GB encoder for the same prompt again. Its sky-high click-through in search tells you people are actively hunting for exactly this relief.

How it works

It's an output node - the end of a branch, not a passthrough. You feed it the CONDITIONING from a Gemma text encode, and it writes that encoded tensor to a file. Later, the companion LTXV Load Conditioning node reads it back and hands it straight to the model, skipping the encoder entirely. Lightricks built this save/reuse pair specifically to kill the constant loading and unloading of the Gemma encoder around every generation. If your workflow reuses prompts - a fixed style prompt, a batch with shared conditioning, or just heavy iteration on one idea - this turns a repeated 22GB penalty into a one-time cost.

The inputs

• conditioning - the encoded prompt to save.
• filename (default "conditioning") - what to call it on disk. Name it something you'll recognize later; the Load node picks from these by name.
• dtype - the storage precision. Lower precision means smaller files; match it to how you'll use the conditioning so you're not upcasting garbage later.

Because it's an output node it has no outputs - it terminates the branch and writes the file as a side effect.

Installing it

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

Common issues

Nothing seems to happen when you run it. That's expected - it's an output node with no visible output. Check your ComfyUI output/conditioning directory for the file. If it's there, it worked.

The Load node can't find your file. Names have to line up. Whatever you set in filename is what shows up in Load Conditioning's dropdown; a mismatch means you're looking for the wrong thing. Keep the names descriptive and distinct.

Your saved conditioning is "wrong" later. A saved encoding is frozen at the moment you saved it - prompt, encoder, and all. Change the prompt and the cached file is stale; re-encode and re-save. Cache the prompts you've finished tuning, not the ones still in flux, and you'll never get bitten by this.