🔥 Winnougan LTX Load Conditioning

Once you've saved a conditioning file with Winnougan LTX Save Conditioning, this is the node that makes the habit pay off: it loads your cached Gemma text encoding from disk in milliseconds, no text encoder in memory, no VRAM spent, no wait. It's the "second run and every run after" half of the pair - the one that turns a 30-second iteration loop into a near-instant one.

The idea is simple and it's the same caching instinct you already use for VAE or model loads: the expensive thing (encoding a prompt with Gemma) only needs to happen when the prompt changes. Iterating on seed, resolution, sampler, or NAG strength? The encoding is byte-identical every time. Drop the text encoder from the graph entirely and feed this node's output where the encoder's CONDITIONING used to go.

How it works

The dropdown lists every .safetensors file in your ComfyUI/models/embeddings folder. On execution, the node reads the file, rebuilds the conditioning list from the saved tensors and attention masks, and reconstructs a proper CONDITIONING output - including metadata like the saved dtype and timestamp shown in the load_info string. It even watches the file's modification time via IS_CHANGED, so if you re-save over the same filename it re-runs instead of serving you a stale cached result.

Inputs

• file_name - dropdown of saved conditioning files. Shows "(no files found)" if the folder is empty, which means you haven't run the Save node yet.
• device - cpu is the default and the safe choice: the conditioning loads into system RAM, works on any setup, and the sampler pulls it over as needed. gpu loads it straight to VRAM, marginally faster, at the cost of VRAM you might not have to spare.

Outputs are conditioning (wire it exactly where the text encoder's output went) and load_info (a status string for a Show Text node).

Install

It's part of the ComfyUI_WLTX_nodes pack - install once, get all eight nodes under Winnougan LTX:

cd ComfyUI/custom_nodes
git clone https://github.com/Winnougan/ComfyUI_WLTX_nodes

Restart ComfyUI, or search ComfyUI_WLTX_nodes in ComfyUI Manager. No pip dependencies beyond what ComfyUI already ships.

Where people get burned

The classic mistake is wiring Load in alongside the text encoder and expecting ComfyUI to prefer one - you have to actually remove the encoder (or bypass it) or you're still paying the Gemma cost. Second: loading on gpu when you're VRAM-tight can be the thing that tips an otherwise-fine workflow into an OOM; the CPU default is genuinely fine here, so leave it unless you're chasing milliseconds. And a reminder that applies to the whole pack: these nodes are only as good as the workflow around them - a saved file from a different prompt is still a valid file, just not the one you meant.