Wan Video Power Lora Loader (QQ)

If you've built a Wan 2.1 or 2.2 workflow in Kijai's ComfyUI-WanVideoWrapper, you know the LoRA situation gets ugly fast. One node per LoRA, each with its own model-and-clip in and out, and pretty soon half your graph is LoRA daisy-chains. WanVideoPowerLoraLoader is the pack's answer: one node with an rgthree-style dynamic UI that stacks as many LoRAs as you want, then hands the whole thing to the wrapper in a form it actually understands.

It's the "v1" loader in this pack, and its defining trait is that it doesn't return a patched model at all. Its two outputs are selected_lora and low_lora, both of type WANVIDLORA - a LoRA spec rather than a model. That's the format Kijai's wrapper consumes, so the intended flow is: build your stack here, wire the selected_lora output into the wrapper's WanVideoSetLoRAs node, and the wrapper applies everything in one place. This is also where the Wan 2.2 high/low split shows up: low_lora is the same stack earmarked for the low-noise pass, so you can keep the high-noise pass clean and let the low side carry speed LoRAs at lower strength.

The inputs

There are only two, and both are optional:

• prev_lora (WANVIDLORA) - for chaining multiple loaders. Feed the selected_lora output of one loader into the next and they accumulate instead of overwriting.
• blocks (SELECTEDBLOCKS) - per-LoRA layer/block selection, for the people who want to limit which transformer blocks a LoRA touches (handy for style LoRAs that should leave structure alone).

The actual LoRA list, strengths, and the high/low pair detection all live in the node's dynamic UI - that's the rgthree-style interface where you add rows, toggle pairs, and set dual High/Low strength sliders. The "Lora high.low filter" circle icon cycles through high-only / low-only / both.

Choosing between this and PowerLoraLoaderV2

This pack ships both, and the difference is architectural. WanVideoPowerLoraLoader outputs WANVIDLORA objects meant for WanVideoSetLoRAs inside the KJ wrapper graph - the right choice when you're already running wrapper-native pipelines (with WanVideoLoraSelect, block swap, TTM, all the wrapper machinery). PowerLoraLoaderV2 instead outputs patched MODEL/MODEL_LOW/CLIP you can wire straight into a sampler, fewer moving parts but less wrapper-native. If a downloaded workflow already uses WanVideoSetLoRAs, this is the loader it expects.

Installing it

It ships in siraxe's community pack (the "QQ" extension layer for the Wan wrapper). Install via ComfyUI Manager by searching WanVideoWrapper_QQ (or SA-Nodes-QQ), or:

cd ComfyUI/custom_nodes
git clone https://github.com/siraxe/ComfyUI-WanVideoWrapper_QQ.git

Restart ComfyUI. You also need Kijai's ComfyUI-WanVideoWrapper installed - this node is pointless without it, since WANVIDLORA is the wrapper's type. Your LoRAs go in ComfyUI/models/loras.

Gotchas

The repo was renamed to ComfyUI-SA-Nodes-QQ in v1.3.4, and old workflows referencing the old wanwrapper_qq pack can silently install the wrong thing - delete any stale wanwrapper_qq folder in custom_nodes. Also remember the pack is explicitly experimental: the README's own changelog is full of WIP caveats, so if a version update changes the UI mid-project, that's the expected ride. And one Wan-specific warning that applies to every LoRA loader here: resist cranking speed LoRAs on the high-noise pass - that's the fastest way to get "flux-level plastic skin" complaints from your own outputs.