🅛🅣🅧 LTXV Load Conditioning

This little node solves a very specific, very real headache: the LTX-2 text encoder is a 22GB monster that has to load, run, and unload every time you tweak something unrelated and re-run the graph. If your prompt hasn't changed, that's pure wasted time and VRAM churn. Load Conditioning lets you skip it - read a previously-encoded conditioning straight off disk and feed it to the model, no Gemma in sight.

How it works

It's the read half of a save/load pair. You encode a prompt once (with the Gemma nodes) and write the result to disk with LTXV Save Conditioning; this node loads that file back as a CONDITIONING object. From then on, iterating on samplers, guiders, seeds, or resolution costs you nothing on the text-encode side, because the encoding is already done and cached. Lightricks shipped these save-and-reuse nodes specifically to attack the constant loading and unloading of the 22GB encoder, which was the worst LTX-2 pain point on consumer cards. If you're tuning a workflow around a fixed prompt, this is one of the highest-leverage nodes in the pack.

The inputs and output

Two dropdowns, both simple:

• file_name - the saved conditioning file to load, picked from what's on disk.
• device - where to place the loaded tensor (your GPU or CPU). Put it where the model expects it; CPU if you're squeezing VRAM.

The single output is CONDITIONING, which drops in anywhere a text-encode's output would go - sampler, guider, or a conditioning-combine node. As far as the rest of the graph is concerned, it's indistinguishable from a freshly-encoded prompt; it just arrived without the 22GB detour.

Installing it

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

No model download - it just reads files you saved earlier.

Common issues

The file dropdown is empty. You haven't saved anything yet. This node has nothing to load until LTXV Save Conditioning has written a file - they're a set, and the save node comes first.

Device mismatch errors. If the loaded conditioning lands on the wrong device for your model, you'll get a tensor-device error. Set device to match where the model runs; the whole point of these encoders living on separate devices is to manage VRAM, so pick deliberately.

You changed the prompt and forgot. The obvious footgun - a cached encoding is frozen. If you edit your prompt, you have to re-encode and re-save, or you're generating from stale conditioning. Cache the encodings you're done iterating on, not the ones you're still writing.