Load WanVideoLora From HF

Like its CivitAI- and ComfyOnline-sourced siblings in this pack, this node produces a WANVIDLORA object for Kijai's ComfyUI-WanVideoWrapper to consume - it isn't a general MODEL/CLIP LoRA loader, and stock ComfyUI doesn't know what to do with its output on its own. Install WanVideoWrapper first if you don't already have it; this node is a drop-in for its own LoRA-select node, just sourced from a HuggingFace repo instead of a local file.

What it's for

A good chunk of Wan's LoRA and speed-adaptation ecosystem - the LightX2V and CausVid step-distillation LoRAs among them - gets published or mirrored on HuggingFace, sometimes ahead of a CivitAI upload. This node fetches a named file from a public repo directly at run time rather than requiring it already downloaded into your LoRA folder, which is the difference that matters when you're running ComfyUI somewhere without persistent, pre-populated storage - a cloud executor or API worker being the typical case.

The inputs and outputs that matter

• repo (STRING, required) - the HuggingFace repository name in user/repo form, per the node's own tooltip.
• file_name (STRING, required) - the exact filename inside that repo, matching path and extension.
• strength (FLOAT, default 1, range -10 to 10, step 0.0001, required) - "LORA strength, set to 0.0 to unmerge the LORA." Worth knowing the community-tuned range for Wan 2.2's speed LoRAs specifically: 0.4–0.6 on the low-noise denoising pass, not full strength across both passes - running them hot is a well-documented way to strip out motion, lighting and composition quality in exchange for speed you may not even need.
• prev_lora (WANVIDLORA, optional) - chains multiple LoRAs together: feed one node's lora output into the next one's prev_lora, then wire the final result into WanVideoWrapper's model loader.
• blocks (SELECTEDBLOCKS, optional) - restricts the LoRA to specific transformer blocks rather than the whole model. Advanced, rarely needed; comes from a block-selection node elsewhere in WanVideoWrapper.
• Output: lora (WANVIDLORA) - into the next node's prev_lora, or straight into the wrapper's model loader.

One gap to know about: there's no token-carrying variant of this node, unlike the plain LoRA loader elsewhere in this pack (LoadLoraFromHF has an HFWithToken sibling). If the Wan LoRA you want sits in a private or gated HuggingFace repo, this node can't reach it - public repos only.

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 the output to connect to anything. Neither pack bundles Wan's base-model weights, which are a much larger, separate download through your normal Wan setup.

Common issues & troubleshooting

404 or repo-not-found. Verify repo is exactly user/repo with no extra path, and that file_name matches the file's real location and extension inside that repo.

It's a gated or private repo. No workaround here - this node has no auth field, so it's limited to public repos.

Nowhere to plug the lora output. Install ComfyUI-WanVideoWrapper - WANVIDLORA is that pack's own type.

Motion goes flat, faces lose consistency. Almost always a speed LoRA run too hot. Back it off to the 0.4–0.6 range on the low-noise pass rather than assuming the default of 1.0 is safe for this category of LoRA specifically.

LoRA has no effect 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 anything trained for an unrelated architecture loads silently and does nothing.