LivePortrait Load FaceAlignmentCropper

The newest detector in the pack, added in a later update to fix the exact gap the other two leave open: small faces. LivePortraitLoadFaceAlignmentCropper loads a detector built on the face-alignment library (1adrianb's), which can use two blazeface variants and the SFD detector - and crucially, blazeface's back-camera model detects faces much farther away than the short-range blazeface that MediaPipe is stuck with. If you've ever run a wide shot through LivePortrait and gotten a row of black frames where the face should be, this is the node that fixes it.

The detector choice

• blazeface_back_camera (default) - the model tuned for selfie-camera and rear-camera distances, i.e. it finds small faces at distance. This is the whole reason to use this node.
• blazeface - the short-range front-camera model; about on par with what MediaPipe offers.
• sfd - the older Single Shot Face Detector. Different failure profile, sometimes better in odd lighting.

Two things worth knowing before you switch:

1. The first warmup is slow. The README warns the initial run can take a long time while it loads. Subsequent runs are quick. Don't assume it's hung - let it finish once.
2. The license story stays clean. This detector path doesn't pull in InsightFace's non-commercial weights, so you keep the permissive stack.

The device controls

Three device/dtype inputs to get right:

• landmarkrunner_device - where the landmark model runs; includes torch_gpu like the MediaPipe node (the PyTorch-converted landmark runner, no ONNX needed).
• face_detector_device (cuda / cpu / mps) - where the detector itself runs. Default cuda; on Apple silicon pick mps.
• face_detector_dtype (fp16 / bf16 / fp32, default fp16) - precision for the detector. fp16 is fast and fine; drop to fp32 if you see detection weirdness.

Plus keep_model_loaded (default on). Output is a cropper (LPCROPPER) into LivePortrait Cropper - same hand-off as the other two loaders, so swapping detectors in an existing workflow is a one-node change.

Where it fits

The decision tree for the three loaders: InsightFace when you need extreme-angle robustness and don't care about the license, MediaPipe as the clean default, and this one when MediaPipe can't find the face - small faces, wide shots, faces far from the driver. It's the least-tested of the three, so expect to fiddle with face_detector_dtype or switch to sfd if your source confuses it. And remember that warmup cost is one-time; after that it behaves like a normal detector.