Load Wan 2.2 Lora

This is LoadWan2_2FunLora's twin for the non-Fun side of the pack - same two-LoRA-slot design, same reason for it, different model family underneath. If you loaded your checkpoint through LoadWan2_2Model (plain Wan 2.2, not the PAI-retrained Fun variant), this is the LoRA node that pairs with it. Wan 2.2's architecture splits denoising across a high-noise expert (handles motion and composition) and a low-noise expert (handles detail), which is exactly why the schema gives you two separate LoRA slots instead of one - you can target either pass independently.

How it works

Feed it a funmodels bundle, assign a LoRA to the low-noise slot, optionally another to the high-noise slot, set a strength, and it returns the patched model. This is where the community's standard advice for Wan 2.2 speed LoRAs - apply to low-noise only, leave high-noise clean - actually gets wired up if you're following that pattern.

The inputs that matter

• funmodels (required) - from LoadWan2_2Model.
• lora_name - LoRA for the low-noise/detail expert.
• lora_high_name - LoRA for the high-noise/motion expert. Leave it unset to only touch the low-noise pass.
• strength_model - default 1, range -100 to 100, applies to both slots.
• lora_cache - default false; enable to skip re-merging weights on repeated runs when the LoRA and strength aren't changing between generations.

Output: funmodels, patched, into Wan2_2I2VSampler.

Installing it

Ships with VideoX-Fun, no separate install step. Search ComfyUI Manager for VideoX-Fun; if it doesn't turn up there:

cd ComfyUI/custom_nodes
git clone https://github.com/aigc-apps/VideoX-Fun
pip install -r VideoX-Fun/requirements.txt

Restart ComfyUI.

Common issues

Wondering whether to put a speed LoRA on both slots. Don't, by default. The strongest, most-repeated warning across the Wan 2.2 community is that step-distillation LoRAs (lightx2v-style, CausVid-style) applied to the high-noise pass wreck scene composition, lighting, and emotional expressiveness - the exact things Wan 2.2 got praised for over 2.1. The workaround people settled on: speed LoRA on low-noise only, keep high-noise clean and full-quality. That's precisely what leaving lora_high_name blank gets you here.

LoRA has basically no visible effect. Check strength_model isn't near zero, confirm the LoRA was actually trained for Wan (2.1-trained Wan LoRAs are broadly forward-compatible with 2.2, per community reports, but LoRAs for other architectures aren't), and make sure lora_cache isn't holding onto a stale merge from a previous run if you changed the LoRA file mid-session.

Wrong node for your checkpoint. This one's for plain Wan 2.2 loaded via LoadWan2_2Model. If your model came from LoadWan2_2FunModel instead, use LoadWan2_2FunLora - the model bundles aren't drop-in interchangeable between the two families even though they share the same FunModels type name.

Empty LoRA dropdowns. Files need to actually be in ComfyUI's loras/ directory, and the list needs a ComfyUI restart to pick up anything added after startup.