WanAnimatePlus SetLoRAs

The author's own description says it plainly: "Sets the LoRA weights to be used directly in linear layers of the model, this does NOT merge LoRAs." That last clause is the whole point of this node, and it's a distinction that saves you a lot of pain.

When you merge a LoRA into a checkpoint, you get a new set of weights - and if you want a different strength, or a different LoRA, you're re-merging. When you set a LoRA directly into the linear layers at load time, the base weights stay untouched and the LoRA's low-rank adapters are applied on the fly. Strength changes are instant, stacking is painless, and the base model file on disk never changes. For a 14B video model, where a merge is a slow, VRAM-hungry operation, that's not a convenience - it's the difference between iterating on styles and waiting out a bake each time.

How to use it

It's deliberately minimal: a model (WANVIDEOMODEL) goes in, an optional lora (WANVIDLORA) goes in, and a model comes out. The WANVIDLORA comes from WanAnimatePlus LoraSelect or LoraSelectMulti, which is where you actually pick the .safetensors file and set the weight. Wire:

ModelLoader → LoraSelectMulti → SetLoRAs → (sampler)

Or for several LoRAs in sequence, chain SetLoRAs nodes or use the multi-select so the stack applies in one shot. Because nothing is merged, the node clones the model patcher and applies the adapters - undo is as easy as disconnecting the wire, which is not something you can say about a merged checkpoint.

A few practical notes from the Wan world: the pack's own docs treat LoRAs as optional, and character or style LoRAs trained for Wan 2.1 carry forward to 2.2, so the library is deep. If you're running speed LoRAs (lightx2v / CausVid), remember the community rule - strength matters as much as presence, and at CFG 1.0 with 4 steps a full-strength LoRA can flatten the output. Keep a strength dial somewhere in reach.

The one trap

SetLoRAs is not where you load the LoRA - it only applies one that's already been selected. Beginners wire a file path in here and wonder why nothing happened. The path lives on the LoraSelect node. Also, TorchCompileSettings has a related flag (allow_unmerged_lora_compile) that gets fussy with some dynamic LoRAs if you compile the whole graph - leave it off unless you know you need it.

Install

Standard for this pack:

cd ComfyUI/custom_nodes
git clone https://github.com/wuwukaka/ComfyUI-WanAnimatePlus.git

Restart ComfyUI, or install via Manager by searching "ComfyUI-WanAnimatePlus". Keep the original ComfyUI-WanVideoWrapper installed, and remember the pack-wide rule: stick to WanAnimatePlus nodes for the whole chain rather than mixing in original WanVideoWrapper ones.