WanVideo Lora Select (LoraManager)

If you run Wan video through Kijai's ComfyUI-WanVideoWrapper - the primary third-party Wan wrapper, the one with SageAttention and TeaCache support - its LoRA plumbing uses a special WANVIDLORA type that the normal LoRA loaders don't speak. This node is the bridge: it lets you pick LoRAs from LoRA Manager's library, with the same autocomplete and trigger-word tracking you use everywhere else, and outputs them in the WANVIDLORA format the WanVideoWrapper model loader expects.

That matters because Wan is where a lot of the LoRA action is right now - speed LoRAs like LightX2V and CausVid that cut generation to a handful of steps, plus a growing library of style and motion LoRAs. Managing those in the same place as your image LoRAs, instead of a separate dropdown pile, is the whole pitch.

There's a sibling node, WanVideo Lora Select From Text, that does the same thing but takes the LoRA list as a plain string input and adds block-level control. Reach for this one when you're picking LoRAs by hand in the node; reach for that one when the selection needs to come from elsewhere in the graph or you want to target specific model blocks.

How it works

You choose LoRAs in the autocomplete box (the same <lora:name:strength> syntax as the rest of the pack), and the node builds a WANVIDLORA object to hand to the WanVideoWrapper. Two switches control how the LoRAs get into the model: merge_loras and low_mem_load. It also passes through the trigger words and a list of what's active, same as the image stackers.

Inputs and outputs that matter

• text (autocomplete LoRA box, required) - your Wan LoRA selection.
• merge_loras (BOOLEAN, default true) - merge the LoRA weights into the model rather than keeping them applied separately. On is the usual default.
• low_mem_load (BOOLEAN, default false) - a lower-VRAM loading path; turn it on if you're memory-constrained.

Outputs: lora (WANVIDLORA - into the WanVideoWrapper's model loader), trigger_words (STRING) and active_loras (STRING).

How to install it

Via ComfyUI Manager: search lora-manager, Install, restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/willmiao/ComfyUI-Lora-Manager.git
cd ComfyUI-Lora-Manager
pip install -r requirements.txt

then restart. Note this node's output only means something if you also have Kijai's ComfyUI-WanVideoWrapper installed - that's a separate pack, and it's what consumes the WANVIDLORA.

Common issues & troubleshooting

The lora output won't connect to anything. You need ComfyUI-WanVideoWrapper installed; the WANVIDLORA type plugs into its model loader specifically. Without the wrapper, nothing accepts that output.

A Wan 2.1 LoRA behaves oddly on Wan 2.2. Transfer between Wan versions is partial - Kijai's own line is that lightx2v speed LoRAs work "to some extent," and some users find likeness gets worse once LoRAs are stacked across versions. Don't assume clean carry-over.

Speed LoRA killed my negative prompt. That's expected: distillation/speed LoRAs force CFG 1.0, where the negative prompt has nothing to act through. It's a property of the LoRA, not this node.