Load WanVideoLora From CivitAI

Before you drop this into a graph: it only means anything alongside Kijai's ComfyUI-WanVideoWrapper. The output type is WANVIDLORA, a custom type that wrapper pack defines and its own model-loader node consumes - this node doesn't work with native ComfyUI's WanVideo support or a plain LoraLoader. If you don't already have WanVideoWrapper installed, install that first; this node is a drop-in alternative to its own LoRA-select node, just sourced from a CivitAI ID instead of a local file dropdown.

What it's for

Community Wan workflows almost always run through WanVideoWrapper rather than ComfyUI's bare-metal Wan nodes, because that's where the ecosystem's LoRA support, block-selective loading and speed tricks live. Its LoRA selector normally points at a file already sitting in your LoRA folder. This node swaps that for a CivitAI model ID, so the fetch happens as part of the run - useful the moment you're not on a machine with your Wan LoRA collection already downloaded, which is most API and cloud-executor setups.

How it works

It resolves the CivitAI ID, downloads the LoRA file, and packages it as a WANVIDLORA object rather than patching a MODEL socket directly - that packaging is what lets WanVideoWrapper's model loader apply it, chain several LoRAs together, and optionally restrict it to specific transformer blocks.

The inputs and outputs that matter

civitai_model_id (STRING, required) - "The ID of the model to download from CivitAI." As with the other CivitAI-sourced loaders in this pack, this needs to be the model-version ID from CivitAI's API, not the plain model-page ID in the browser URL - grab it from the specific version you want, not the model's overview page.
strength (FLOAT, default 1, range -10 to 10, step 0.0001, required) - per the node's tooltip, "LORA strength, set to 0.0 to unmerge the LORA." If this is a speed/distillation LoRA (LightX2V, CausVid and similar), the community-tuned range is 0.4–0.6 rather than 1.0, applied on the low-noise denoising pass specifically - running speed LoRAs at full strength on both passes is a well-documented way to flatten Wan 2.2's motion, lighting and composition, not just make it faster.
prev_lora (WANVIDLORA, optional) - "For loading multiple LoRAs." Chain several of these nodes by feeding each one's output into the next one's prev_lora, then wire the last one into WanVideoWrapper's model loader. This is the standard stacking pattern for the wrapper, not something unique to this node.
blocks (SELECTEDBLOCKS, optional) - an advanced knob for restricting the LoRA to specific transformer blocks rather than applying it everywhere, fed from a block-selection node elsewhere in WanVideoWrapper. Most people skip it; it's there for when you want surgical control, for example keeping a motion LoRA off the blocks that carry identity.
Output: lora (WANVIDLORA) - feeds into prev_lora on the next LoRA node in the chain, or directly into WanVideoWrapper's model loader if it's the only one.

How to install it

ComfyUI Manager - search "comfyui-model-dynamic-loader," install, restart.
Manual - cd ComfyUI/custom_nodes && git clone https://github.com/jax-explorer/comfyui-model-dynamic-loader, then restart.

You also need ComfyUI-WanVideoWrapper installed separately for this node's output to connect to anything - search for it in the Manager the same way. Neither pack's README bundles Wan model weights; those are a separate, much larger download you handle through your normal Wan setup.

Common issues & troubleshooting

Nothing to plug the output into. That's WanVideoWrapper missing, not this node being broken - WANVIDLORA is defined by that pack, so nothing else in a stock ComfyUI install will accept it.

Motion and detail collapse after adding a LoRA. If it's a speed LoRA, you're very likely running it too hot - try 0.4–0.6 on the low-noise pass rather than full strength everywhere; this is a widely reported failure mode on Wan 2.2, not something specific to loading it this way.

CivitAI ID doesn't resolve. Model-page ID versus model-version ID is the usual culprit - see above. It's also worth checking the model hasn't simply been taken down; CivitAI's catalog loses a meaningful share of older uploads over time.

LoRA does nothing at all. Confirm it was actually trained for Wan and for the version you're running - 2.1-trained LoRAs generally carry forward to 2.2, but a LoRA built for an unrelated architecture won't error, it'll just have no effect.