☁️BizyAir ReActor 🌌 Face Booster

This node doesn't swap a face - it configures how one gets cleaned up afterward. ReActor is a post-process face swap: it drops a source face onto a generated image at the pixel level, after the fact, which is a fundamentally different move from generative identity tools like IP-Adapter FaceID or InstantID that guide the diffusion process itself toward a likeness. ("FaceID gets you 80% there, ReActor gets you the rest of the way there" is the standard community framing of how the two combine.) This particular node produces a FACE_BOOST config object that feeds into BizyAir's main ReActor swap node - on its own it does nothing visible.

What it configures

boost_model picks which face-restoration model runs on the swapped face afterward: none, codeformer-v0.1.0.pth, GFPGANv1.4.onnx, GPEN-BFR-2048.onnx, or RestoreFormer_PP.onnx. These are the standard face-restoration models you'll recognize from A1111's "Restore faces" toggle or ADetailer - CodeFormer and GFPGAN are the two everyone's heard of; GPEN and RestoreFormer are less common alternatives with their own tuning.

codeformer_weight (0–1, default 0.5) only matters when boost_model is CodeFormer - it's CodeFormer's own fidelity-versus-quality dial. Lower values lean toward heavier restoration (smoother, less faithful to the swapped face's actual texture); higher values stay closer to the raw swap output. visibility (0.1–1, default 1) is the overall blend strength of the boosted result against the unboosted swap. interpolation (Nearest/Bilinear/Bicubic/Lanczos, default Bicubic) is the resampling method used when the boost model's output gets resized back into the frame - Bicubic is a sane default and rarely worth touching.

restore_with_main_after (default off) runs the main ReActor node's own restoration pass a second time, after this boost - useful in specific cases, but stacking two restoration passes is also how you get over-smoothed, plasticky-looking faces, so it's not a "turn it on for extra quality" switch. enabled is the master on/off for the whole boost step.

Installing it

ComfyUI Manager: search BizyAir, install, restart. Manual:

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

Restart after. This node's config doesn't require a local download by itself, but the ReActor pipeline it feeds into runs on InsightFace's ArcFace recognition backbone, which historically has been one of the more painful installs in local generation (C++ build tooling, platform-specific wheels) - worth budgeting time for the main ReActor node's setup even if this particular config node is trivial.

Where people get burned

Stacking restoration passes. The classic complaint about face-restore models generally is that they conflict when layered - running this boost and then a separate restore step elsewhere in your workflow (or restore_with_main_after on top of an already-boosted result) tends to produce over-sharpened, waxy faces rather than better ones. Pick one restoration pass and tune it, don't chain several.

Setting codeformer_weight and picking a different boost_model. That value is inert unless boost_model is actually set to CodeFormer - if you're on GFPGAN or GPEN, it does nothing.

Treating this as the swap node. If you dropped this into a graph expecting a face to swap, you're missing ReActor's actual swap node - this is purely the post-swap cleanup config.