🅛🅣🅧 LTXVQ8Lora Model Loader

This is a LoRA loader specialized for LTX running in Q8. If you're already using the quantized LTX path to fit the model in less VRAM, you can't just slap a normal LoRA loader on top - the LoRA has to be applied in a way that plays nicely with the Q8 weights. LTXVQ8LoraModelLoader is the node that does that: pick a LoRA, set a strength, and it patches it onto your (Q8) LTX model correctly.

For context on why Q8 exists in LTX-land at all: quantization is how these big video models fit on consumer cards, and Q8 is one of the tighter, higher-quality quantization levels. The KB's history here is worth knowing - the old LTX-Video Q8 kernels were genuinely painful to install ("the most confuse installation of my life," one user; another gave up after three hours), and they don't provide a speedup on RTX 30-series cards, only help you squeeze into memory. A loader that handles Q8 LoRA application for you is a direct response to that friction.

How it works

It takes your loaded MODEL (the Q8 LTX model), applies the selected LoRA at the strength you set, and returns the patched MODEL. Chain it after your model/quantized loader and before your sampler, exactly where any LoRA loader sits - the difference is that the merge is Q8-aware so you don't corrupt the quantized weights.

The inputs and outputs that matter

• model (MODEL) - your LTX model, in.
• lora_name (dropdown) - the LoRA to load. This list is populated from your loras folder, so if it's empty, that's why (see below).
• strength_model (FLOAT, default 1) - how strongly the LoRA applies. 1.0 is full; dial down for a subtler effect. Note the KB's tip for LTX-2.3 two-stage distilled workflows: LoRA strengths like 0.4 on the first sampler and 0.5 on the second are common starting points, so don't assume 1.0 is always right.

Output: a single MODEL, into your sampler (or the next LoRA loader if you're stacking).

How to install it

ComfyUI Manager: Ctrl+M, Install Custom Nodes, search LTXVideo, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Lightricks/ComfyUI-LTXVideo

then restart. Official Lightricks pack. You supply the actual LoRA files - drop them in ComfyUI/models/loras. The pack ships a lot of official LoRAs (camera-control, IC-LoRAs, the distilled LoRA), and remember the distilled LoRA alone is ~7GB, so budget disk. Plus the usual LTX-2 stack (32GB+ VRAM, 100GB+ disk).

Common issues & troubleshooting

The lora_name dropdown is empty. No LoRA files in your models/loras folder, or ComfyUI didn't rescan. Download the LoRA you want (the README links every official one), drop it in models/loras, and restart ComfyUI so it repopulates the list.

The LoRA does nothing / barely anything. Check strength_model isn't near zero, and confirm the LoRA actually matches your base model generation. A LoRA trained for a different LTX version can load without erroring and just not do much.

It cut to something different after the first frame. The KB notes exactly this when loading full-model LoRAs on the distilled model - the fix people found was bumping the step count (one user set steps to 12 to solve it). Try more steps before assuming the LoRA is broken.

You expected Q8 to be faster. On RTX 30-series it generally isn't - Q8 in LTX is about fitting into memory, not speed, per Lightricks' own comment. If you're chasing speed, that comes from the distilled checkpoint and low step counts, not from quantization.