Load Lynx Resampler

This is the loader for the small-but-crucial piece of ByteDance's Lynx identity system: the resampler. Lynx keeps a person's face consistent across a Wan video by turning an ArcFace face embedding into a compact set of identity tokens the model can condition on - and the thing that does that conversion is a Perceiver Resampler. LoadLynxResampler puts that adapter into memory and hands it downstream as a LYNXRESAMPLER. It's a plumbing/loader node, but without it the Lynx identity path has no brain.

If you're setting up face-consistent Wan generation with Lynx, this is one of the first nodes you place, alongside the face-crop node and the main Lynx embeds node.

How it works

The resampler is a lightweight adapter (not a full model) that takes the raw face embedding - a vector describing facial geometry, produced by InsightFace's ArcFace - and resamples it into the token format Wan's attention layers accept. This node loads that adapter's weights at the precision you choose and exposes it as an object the rest of the Lynx pipeline consumes. All the heavy lifting is elsewhere; this just gets the adapter ready.

The inputs and outputs that matter

• model_name (enum) - which resampler file to load. The dropdown is populated from your local models folder, so if it's empty, you haven't downloaded the Lynx resampler weights yet - that's the number-one gotcha here.
• precision (fp16 default, or bf16 / fp32) - load precision. fp16 is the standard choice and keeps it light; go bf16/fp32 only if you hit numerical oddities.

Output is resampler (LYNXRESAMPLER) - feed it into the Lynx pipeline that produces the identity embeds (which ultimately reach WanVideoAddLynxEmbeds).

How to install it

Comes with the WanVideoWrapper. ComfyUI Manager: search ComfyUI-WanVideoWrapper, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-WanVideoWrapper
pip install -r ComfyUI-WanVideoWrapper/requirements.txt

then restart. The node code is one thing; the resampler weights are a separate download you drop into the appropriate models subfolder so the model_name dropdown has something to show. Lynx as a whole also depends on InsightFace for the face embedding - the single fiddliest install in the identity-tool world, though its 1.0 release finally dropped the C++ build requirement, so a fresh install is less painful than the horror stories suggest.

Common issues & troubleshooting

The dropdown is empty. No resampler file found. Download the Lynx resampler weights into your models folder and refresh. This is by far the most common reason people land on this page confused.

It loads but identity still doesn't work. The resampler is only one leg of Lynx. You also need InsightFace producing a real ArcFace embedding (via the face-crop node) and the main WanVideoAddLynxEmbeds node wiring it into the sampler. A loaded resampler with no upstream face is doing nothing.

Precision-related errors. If fp16 throws something numerical on your setup, try bf16. It's a small adapter, so the memory difference is negligible.