LTXV LoRA Selector

The LTXV LoRA Selector is the picker half of this pack's two-node trick. It doesn't apply anything to anything - no weights get touched here. What it does is turn a LoRA file from your models/loras folder into a small descriptor (path, name, strength, and any block selection) that the LTXV LoRA Loader later merges into the model. Think shopping cart, not checkout.

The reason you'd reach for it over ComfyUI's built-in LoraLoader is the chain. Feed one selector's lora output into another selector's previous_lora input and you can stack any number of LoRAs, each with its own strength, then hand the whole pile to a single Loader at the end. The official LTXV workflow had no clean equivalent for that - stacking always meant "unmerge this one, apply that one, redo it in the next sampler," which is the fiddliest part of LoRA work in any model.

The inputs that matter

• lora - a dropdown populated from ComfyUI/models/loras. The author's own tooltip is the spec: it expects .safetensors files. Drop the file in that folder and refresh the node.
• strength (FLOAT, default 1.0, range -10 to 10) - how hard the LoRA pushes. The tooltip gives you the useful edge case: set it to 0.0 to unmerge. Negative strengths invert a LoRA's effect, which is occasionally exactly what you want.
• previous_lora (LTXVLORA, optional) - the chain link. When stacking, wire the previous selector's output here.
• blocks (SELECTEDBLOCKS, optional) - accepts the output of the pack's LTXV LoRA Block Edit node if you want to restrict the LoRA to specific transformer blocks. Skip it unless you're debugging.

The single output, lora, goes to the LTXV LoRA Loader - or into the next selector.

Strength discipline

LTX LoRAs want far less than the SDXL instinct of 1.0. Community testing on the 2.3-era distilled LoRA settled around 0.4–0.5, and cranking strength is a leading cause of the "diseased" skin and melted faces that fill the help threads. Start at or below 0.5, verify the effect, then bump. When something looks wrong and you can't diagnose it, set strength to 0 - the unmerge trick - to confirm the LoRA is the culprit before you chase anything else.

Installing it

ComfyUI Manager: search ComfyUI-LTXVideoLoRA. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/dorpxam/ComfyUI-LTXVideoLoRA

Restart, and remember the selector is not a loader - installing the pack alone does nothing until a Loader sits at the end of the chain. If the repo 404s (it has been intermittently), Manager's registry is the reliable route.

What trips people up

• The dropdown reads ComfyUI/models/loras specifically, and only when the node refreshes. New files need a refresh.
• Chain order matters. The Loader applies LoRAs in the order the selectors were linked, and order shifts results when two LoRAs fight over the same weights.
• The selector can't know what model a LoRA was trained for. It will happily describe a 13B LoRA onto a 2B pipeline, and the mismatch only shows up later as "nothing happened." Check the LoRA's base model before you chain it.