☁️BizyAir ReActor 🌌 Fast Face Swap

ReActor is the other branch of face identity work - not the "generate a face that resembles this person" approach PuLID and InstantID take, but a straight pixel-level swap: detect the face in a target image, detect a face in a source image, replace one with the other. It's post-processing, not generation. The community's own shorthand for the split: adapters like PuLID get you "80% there," and a ReActor-style swap on top gets you the rest of the way to an exact match - the two approaches are complementary rather than competing.

What's actually running underneath

swap_model offers inswapper_128.onnx, and that name is worth reading literally: the open swap model this whole category of tool depends on outputs at 128×128 resolution, full stop, and that ceiling hasn't moved in years - the model's own maintainers describe it as no longer maintained and point commercial users at their paid product instead. That's not a BizyAir limitation, it's the actual state of the open ecosystem this node (and Roop, and FaceFusion, and every other InsightFace-based swapper) is built on. This is exactly why face_restore_model exists in this node: without a restoration pass, a 128×128 swap composited back into a full-resolution image looks visibly soft. Budget for using it.

The inputs that matter

• input_image / source_image - target (the image getting a new face) and source (where the face comes from). source_image is optional in this schema; if you're reusing a face from earlier in a workflow, check whether it's actually wired in before assuming a swap silently failed.
• swap_model, facedetection - the swap model and which detector backbone finds faces to swap (RetinaFace variants or YOLOv5 options).
• face_restore_model - none, codeformer, GFPGAN, GPEN, or RestoreFormer_PP. Pick one; leaving this on none means you keep the raw 128×128-sourced blur. face_restore_visibility (0.1–1) blends the restored face back against the swap, and codeformer_weight (0–1) is Codeformer's own fidelity-vs-quality dial, relevant only if you picked it.
• detect_gender_input / detect_gender_source, input_faces_index / source_faces_index - filters for when a photo has more than one face: restrict by gender, or pick a specific face by its detected position.
• enabled - a toggle to switch the whole node off without unplugging it, handy while you're testing a workflow.

Outputs: SWAPPED_IMAGE (the result), ORIGINAL_IMAGE (an unmodified passthrough, useful for side-by-side comparison), and BIZYAIR_PLACEHOLDER, whose exact purpose isn't documented in the node's own schema - treat it as an internal artifact rather than something to wire up deliberately.

Installing it

ComfyUI Manager: search BizyAir, install, restart - or:

cd ComfyUI/custom_nodes
git clone https://github.com/siliconflow/BizyAir.git

Set your BizyAir API key before running anything.

Where people get burned

Multi-face images are the recurring trap: without setting input_faces_index/source_faces_index deliberately, you're at the mercy of whatever order the detector found faces in, which can silently swap the wrong person in a group photo. And on the licensing side, worth knowing even running through BizyAir's cloud: InsightFace's ArcFace weights underneath this whole swap pipeline are non-commercial-only, and that restriction travels with every tool built on them - switching between ReActor, PuLID, or InstantID doesn't get you a cleaner license, since they all share the same dependency underneath.