Load Wan Lora

LoadWanLora sits between your model loader and your sampler, and its whole job is to patch a LoRA onto whatever funmodels bundle just came out of LoadWanModel. It takes a funmodels in and returns a funmodels out - which is the detail that matters: that in/out symmetry means you can chain several of these in a row to stack multiple LoRAs on the same run, one node per LoRA.

Why you'd want this

Wan's LoRA ecosystem is one of the deeper reasons people stick with it over newer video models - character LoRAs trained on 20-30 still images are the community's standard way to hold a face or outfit consistent across generations, and crucially, LoRAs trained on Wan 2.1 carry forward cleanly to Wan 2.2. If you've trained (or downloaded) a Wan LoRA for a specific look, style, or character, this is the node that applies it.

The fields

• lora_name - the LoRA file to load, picked from a dropdown populated by whatever's sitting in the LoRA folder your VideoX-Fun install scans. If the dropdown is empty, the file isn't where the node expects it - check your install's model layout before assuming the LoRA itself is broken.
• strength_model - how hard the LoRA is applied. The slider technically goes from -100 to 100, but ignore that range; in practice LoRA strengths that actually produce usable output live much closer to 0-1.5. Default is 1.0 (full strength as trained). Push past ~1.2-1.5 and you'll typically start seeing the same kind of overcooked artifacts LoRAs produce in image workflows - burned contrast, exaggerated features, motion getting weird. Negative values invert the LoRA's effect, which is occasionally useful but rarely what you want by accident.
• lora_cache - a boolean, off by default. As the name suggests, enabling it keeps the LoRA's patched weights cached rather than re-merged from disk on every run. If you're queueing multiple generations back-to-back with the same LoRA (batch runs, API usage), this can save you the re-merge overhead each time; for a single one-off generation it won't make a noticeable difference either way.

Output is funmodels - wire it straight into your sampler, or into a second LoadWanLora if you're stacking.

Installing VideoX-Fun

ComfyUI Manager: search "VideoX-Fun" (or "CogVideoX-Fun", the pack's older name, which also surfaces it). Manual: cd ComfyUI/custom_nodes && git clone https://github.com/aigc-apps/VideoX-Fun, then restart ComfyUI.

This is Alibaba PAI's own official node pack for their "Fun" model line - not Kijai's ComfyUI-WanVideoWrapper, which is the more widely used community alternative for running plain Wan. The category you'll see in ComfyUI, CogVideoXFUNWrapper, is a naming fossil from when this pack only covered CogVideoX-Fun, before Wan support got added on top.

Common issues

The dropdown being empty is the most common complaint, and it's almost always a folder-location problem rather than a bug in the node - VideoX-Fun keeps its own model directory layout (ComfyUI/models/Fun_Models/) separate from ComfyUI's standard loras/ folder, so a LoRA that works fine in a native Wan workflow may not be visible here until it's placed where this pack actually looks.

The other one is subtler: because strength_model accepts values up to 100, it's easy to fat-finger a value like 10 instead of 1.0 and get a completely broken generation with no obvious error - if your output suddenly looks like noise or nothing like your prompt at all, check this field before anything else. And if you're chaining multiple LoadWanLora nodes to stack LoRAs, remember each one's strength is independent - stacking three LoRAs at strength 1.0 each compounds, and the combined effect is usually stronger than any single one run alone. Start lower than you think you need to when stacking.