WanVideo Lora Select Multi

This is the node where you load your Wan LoRAs - up to five at once, each with its own strength - and hand the stack to the model loader. For most Wan workflows it's not optional, because the single most common thing people add is a speed LoRA (LightX2V or CausVid), and this is where it goes. Character and style LoRAs go here too.

It outputs a WANVIDLORA that plugs into the lora input of WanVideoModelLoader.

How it works

Each LoRA is a small set of weight adjustments layered onto the base model. This node selects them from ComfyUI/models/loras, applies your per-LoRA strength, and by default merges them into the model weights (merge_loras is true) so there's no per-step patching overhead at generation time. You can chain more than five by feeding a previous stack into prev_lora.

The single most important thing to understand isn't a parameter - it's which pass you apply a speed LoRA to. Wan 2.2 splits denoising between a high-noise expert (motion, composition) and a low-noise expert (detail). The hard-won community rule is: keep the high-noise pass clean and apply speed LoRAs to the low-noise pass only, at reduced strength. Put a speed LoRA on the high-noise pass and you get the notorious "plastic skin," flattened motion, and dead composition the PSA threads keep warning about.

The inputs that matter

• lora_0 … lora_4 - up to five LoRA slots, each picked from models/loras.
• strength_0 … strength_4 - per-LoRA strength. Character LoRAs often want ~1.0; speed LoRAs on the low-noise pass usually want ~0.4-0.6, not full strength.
• prev_lora (optional) - chain another WANVIDLORA in to stack beyond five.
• blocks (optional) - target specific transformer blocks with a LoRA instead of the whole model. Advanced.
• merge_loras - leave on (default) for the fastest generation. Turn off only if you have a specific reason to keep them un-merged (e.g. certain torch.compile setups).

Output: a WANVIDLORA for the model loader.

How to install it

Ships with the pack. Via ComfyUI Manager: search WanVideo Wrapper, install, restart. Manually:

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

then restart. LoRA files go in models/loras.

Common issues & troubleshooting

Everything looks plasticky and the motion died. Textbook speed-LoRA-on-the-high-noise-pass. Move the speed LoRA to the low-noise pass only, drop its strength to 0.4-0.6, and keep the high-noise expert unmodified. This is the Wan quality lesson.

Empty dropdowns. No LoRAs in models/loras, or a rescan is needed. Add files and restart.

A LoRA does nothing / errors. Make sure it's actually a Wan LoRA. LoRAs trained for SDXL/Flux won't apply to Wan's transformer. Note also that Wan 2.1-trained LoRAs are forward-compatible with 2.2 - that's expected and fine.

Speed LoRA at CFG > 1 looks wrong. Speed/distillation LoRAs assume CFG 1.0. If you left CFG at 6 with a LightX2V LoRA loaded, set CFG to 1. (And remember: at CFG 1 the negative prompt stops working - that's the tradeoff, not a bug.)