Load Face Model

Once you've saved a face with InstaSwap's Save Face Model node, this is how you pull it back in. Load Face Model is the quietest node in the pack: one dropdown, one output, zero model inference. But it solves a genuinely annoying problem - it means the reference photo no longer has to sit inside your workflow, or on your disk, for you to keep swapping that face.

Here's the shape of it. You drag in a source photo, run it through InstaSwap Fast Face Swap, and let the FACE_MODEL output feed a Save Face Model node. That writes a small .safetensors file. From then on, any workflow that wants that face just drops in Load Face Model, picks the file from the dropdown, and wires its FACE_MODEL output into the Fast Face Swap node's optional face_model input. The source image is gone from the graph entirely.

The single input that matters

• face_model - an enum dropdown listing every .safetensors in ComfyUI/models/instaswap/faces/. Choose none and the node outputs nothing useful - it's the "bypass" option, so the swap node then requires a source_image instead.

That's it. No other inputs, and the only output is FACE_MODEL.

How it works

The saved "model" isn't a trained network - it's the serialized InsightFace face analysis result: the ArcFace embedding, the 2D/3D landmarks, bounding box, even the detected gender and age. Load Face Model just reads those tensors back into memory with safetensors and hands the face off to the swap node, which is exactly what would have been computed from a source photo anyway. That's why loading is effectively instant, and why the result is identical to using the original image - same embedding, same inswapper ceiling.

Installation

Same pack, no extra install step. It ships with ComfyUI-InstaSwap (ComfyUI Manager → search "ComfyUI-InstaSwap", or clone + python install.py). One gotcha after saving a new model: the dropdown is populated at page load, so refresh your ComfyUI page to see new files appear. And a path warning - the README says saved models land in ComfyUI/models/reactor/faces, but the code actually reads and writes ComfyUI/models/instaswap/faces/. Trust the instaswap folder if you go hunting.

Why bother

Two real reasons. First, workflow hygiene: a saved face model is a self-contained little file, so your graph stays clean and your batch jobs don't depend on a reference image that might get moved or deleted. Second, sharing: you can hand someone a tiny .safetensors with a face embedding instead of a full photo. Worth knowing that this is still the same 128×128 inswapper swap when it hits the main node - the saved model is a convenience, not a quality upgrade.