Skip to content

Face Detect Lite

A quick face finder that runs on the GPU. It returns the same five key points as Face Detect, plus a ready-made mask of the face area.

Face Detect Lite — before / after
In ComfyUI Add Node → ComfyUI-JH-PixelPro → face → Face: Detect Lite (YuNet GPU)
Good for Batches and quick previews; feeding Facial Aligner; limiting a later edit to the face area
Speed About 31 ms on an RTX 5090 versus about 84 ms for MediaPipe on CPU (sample portrait)

When to use it

Use Lite when you want the face position fast and do not need the full 468-point mesh. It uses the YuNet network that ships with Kornia, so it works even if MediaPipe is not installed. It uses the GPU when one is available and falls back to CPU.

Stay with Face Detect and Face Landmarks when you need the dense mesh (warping, reshaping) or better handling of tilted and partly hidden faces.

Detect Lite (YuNet) Face Detect (MediaPipe)
Needs MediaPipe No Yes
conf in bbox_json Real detector score The threshold you set
Face mask output Yes (face box) No
Eye points Eye centres Outer eye corners
Feeds the 468-point nodes No Use Face Landmarks

Both feed the aligner. The eye points differ by roughly 10–20 % of the eye distance, which does not matter for alignment.

Quick start

  1. Connect Load Image to Face: Detect Lite (YuNet GPU).
  2. Connect landmarks_json to the landmarks input of Facial Aligner. It reads the first face.
  3. Optional: use face_mask to limit a later edit to the face area.

Settings

Setting Default What it does Tip
mode single_largest One largest face per image, or the top faces by box size. Portraits: keep the default.
max_faces 1 (1–10) Faces per image in multi_top_k. Group shots: raise it.
confidence_threshold 0.6 (0.1–0.95) Minimum detector score. Raise to drop false hits.
nms_threshold 0.3 (0.1–0.9) Overlap above which duplicate boxes are merged. Lower it if you see duplicate boxes on one face.
detect_size 640 (160–1024) Long side of the copy used for detection. Never enlarges. Keep 640 for portraits. Raise only for group shots with small faces.

Outputs:

Output Type Meaning
landmarks_json STRING Per face: [L-eye, R-eye, nose-tip, L-mouth, R-mouth] in pixels.
bbox_json STRING Per face: x, y, w, h (pixels), conf (real score) and batch_index.
face_count INT Faces returned.
face_mask MASK 1 inside the selected face boxes, 0 elsewhere.

Tips from retouching

  • face_mask is a box, not a face outline. Feather it (for example with Mask Edge Smoother) before using it as an edit mask.
  • A larger detect_size is not always better. At very large values the detector can split or misplace big faces and add false hits on close-ups.
  • A pure black frame can trigger a false detection. Use a real image or a higher threshold.
  • Running a batch of portraits? Lite is the better pick: it is faster and needs no extra install.
Under the hood

The image is shrunk so its long side equals detect_size (smaller images are left alone), run through YuNet via kornia.contrib.FaceDetector, and boxes and points are scaled back to original pixels. Faces are sorted by box area. Images smaller than 64 px are padded with their mean colour for detection only. The detector is cached per device. Kornia downloads the weights once into the torch hub cache; if loading fails, the node raises a clear error. The code lives in core/face_detect_lite.py.

Example workflow

Face Detect Lite example workflow. Copy workflows/sample_portrait.jpg into ComfyUI/input/ first.