LTXV LoRA Loader

Back in early 2025 there was a hole in the LTX Video stack. Lightricks shipped official ComfyUI nodes and an official trainer, but no clean way to load a community-trained LoRA into the model - at least not one trained with a-r-r-o-w's finetrainers, which was the tool everyone was actually using. The LTXV LoRA Loader is the patch for that hole. It takes a diffusion model, takes a LoRA (or a whole chain of them) that a sibling node packaged up, and merges them so the LoRA's style or character actually shows up in your video.

One framing note before the details: this pack targets the LTX Video 0.9.x era - the 2B and 13B weights. If you've moved on to LTX-2 or 2.3, ComfyUI's native nodes load LoRAs all by themselves and you don't need any of this. For the old LTXV weights, this node is the missing middle of the workflow.

Why this node exists

The mechanism explains two of the pack's quirks, so it's worth understanding. When you feed this node an LTXVLORA (the pack's own type), it doesn't just hand the file to ComfyUI's standard loader. It opens each .safetensors, renames every key by prefixing diffusion_model. onto it, and only then calls ComfyUI's own load_lora_for_models to patch the model.

That key rewrite is the whole ballgame. Finetrainers saves LoRAs in the diffusers convention; ComfyUI's LTX model patcher expects the Comfy convention, and that mismatch is exactly why those LoRAs wouldn't load through the official nodes at the time. The output is a plain MODEL, which is the reassuring part - from the sampler's point of view, nothing about your graph changed.

The inputs that matter

• model (MODEL) - the diffusion model you're about to patch.
• ltxv_loader_as_input (BOOLEAN, default false) - this is the one that bites people. If your model comes from Lightricks' official LTXV Loader, flip it to true. The README calls it out in a caution block: the official loader wraps the model in a different internal structure, and leaving the flag off quietly produces output with no LoRA effect at all.
• lora (LTXVLORA, optional) - the packaged LoRA(s) from an LTXV LoRA Selector. Wire that node's lora output here.

The single output, model, goes straight into your sampler exactly like the model output of any other loader.

Installing it

ComfyUI Manager is the reliable path: search for ComfyUI-LTXVideoLoRA and install. Manual clone works too:

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

Then restart ComfyUI. There are no extra Python dependencies and no model downloads - the only files involved are your LTXV checkpoint and the LoRA itself, which lives in ComfyUI/models/loras. One honesty note: the upstream repo page has been intermittently returning 404 as of this writing, and Manager pulls from its own registry, so if a bare git clone fails, Manager is the workaround rather than the backup.

The usual ways this goes wrong

• Flag off. You're using the official LTXV Loader and ltxv_loader_as_input is false. The video renders, but the LoRA is a no-show. Set it true.
• Version mismatch. A LoRA trained on one LTX version applied to another is the classic "works partially or not at all" failure - people have been hitting it with LTX-2 LoRAs since January 2026. Check which base model the LoRA was trained on before you blame this node.
• Overweighting. LTX is not SDXL. Strengths near 1.0 smear faces and skin texture, and LTX users learned to keep weights low - the 2.3-era distilled LoRA, for instance, likes 0.4–0.5. If output looks off, dial the selector's strength down before you suspect the loader.