WanVideo Lora Select

If you're running Wan the way most people do, you have at least one LoRA in the graph - a speed LoRA like lightx2v or CausVid to make generation bearable, maybe a character LoRA on top, maybe a style one. This node is how you feed them into a WanVideoWrapper pipeline. It picks a LoRA file from ComfyUI/models/loras, sets its strength, and - the important part - chains to other LoRA-select nodes so you can stack several without a tangle.

It's a small utility node, but it sits on the critical path of basically every serious Wan workflow, so getting the stacking and the block-targeting right matters.

How it works

Each node selects one LoRA and passes out a WANVIDLORA bundle. To stack, you feed one node's output into the next node's prev_lora input, building a chain that the model loader applies all at once. The blocks input is the advanced lever: it lets you apply a LoRA to a specific subset of the model's blocks rather than the whole thing, which is exactly what the community's Wan 2.2 speed-LoRA best practice needs.

The inputs and outputs that matter

• lora - the file to load, from ComfyUI/models/loras.
• strength (default 1) - how hard it bites. For speed LoRAs the community's tuned range is roughly 0.4-0.6 on the low-noise pass, not a flat 1.0.
• prev_lora (WANVIDLORA, optional) - chain input. Connect the previous LoRA-select here to stack; leave empty for the first in the chain.
• blocks (SELECTEDBLOCKS, optional) - restrict the LoRA to chosen blocks. This is how you apply a speed LoRA to only part of the model.
• merge_loras (default true) and low_mem_load (default false) - merge bakes the LoRA into the weights (faster inference); low-mem load trades speed for a smaller memory spike while loading, useful on tight cards.

Output is lora (WANVIDLORA) → into the next select node, or into your Wan model loader's lora input.

How to install it

ComfyUI Manager - search ComfyUI-WanVideoWrapper, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-WanVideoWrapper
pip install -r ComfyUI-WanVideoWrapper/requirements.txt

then restart. Drop your LoRA files in ComfyUI/models/loras. Kijai's extracted lightx2v/CausVid speed LoRAs live on the Kijai/WanVideo_comfy HuggingFace repo; character LoRAs are whatever you've trained or grabbed.

Common issues & troubleshooting

Speed LoRAs "kill everything that's good" in Wan 2.2. This is the single most-repeated Wan warning, from a PSA thread that generated thousands of test clips: distill LoRAs flatten scene composition, lighting, motion, and give you plastic skin. The community's accepted workaround uses Wan 2.2's two-expert architecture - apply the speed LoRA to the low-noise pass only, keep the high-noise pass clean, and pull strength back to 0.4-0.6. The blocks input is what lets you target it that way.

Your negative prompt stopped working after adding a speed LoRA. Expected. Speed LoRAs force CFG 1.0, and at CFG 1 there's no classifier-free guidance for the negative to act through. That's not this node misbehaving - it's the cost of the LoRA. Reach for a NAG node if you want negative control back.

Empty lora dropdown. It reads ComfyUI/models/loras. New files need a ComfyUI restart (or a refresh) before they appear.