🧑✂️AGSoft Crop Face

Cropping faces by hand is a miserable way to spend an afternoon - until you realize the community's answer is usually a face-detection node you have to install separately, with a model you have to download. 🧑✂️ AGSoft Crop Face sidesteps that: it bundles RetinaFace and its weights inside the pack, so you get solid face detection plus cropping the day you install comfyui-AGSoft. No extra pip install, no hunting for a .pth file.

The short pitch: give it an image (or a batch), it finds the faces, crops them with padding, and outputs both the crop and the bounding-box data. This is the node you reach for when you're building face swap, face restorer, or face-LoRA pipelines and you're tired of drawing boxes.

How it works

Detection runs on the RetinaFace MobileNet0.25 model vendored in the pack's Pytorch_Retinaface/ folder - the weights ship with the repo, so it works offline, first run. Two face-related numbers are worth untangling because they sound identical:

• max_faces_per_image (default 50) - the detection limit. How many faces RetinaFace is allowed to return.
• number_of_faces (default 5) - how many you actually get, starting from start_index in the confidence-ranked list.

So number_of_faces=3 with start_index=1 hands you the 2nd, 3rd, and 4th most confident faces. That's your "skip the blurry one in the back" control.

confidence_threshold (default 0.7) filters weak detections. scale_factor (default 2) is the crop padding - 2 means the crop is twice the face box, so there's context around the face, not a skin-tight rectangle. shift_factor_x/shift_factor_y (both 0.5) shift the crop's center point horizontally/vertically - handy when faces sit low in the frame or you want the crop biased toward one side. aspect_ratio (default 1:1) forces the crop's shape; divisible_by (default 8) rounds crop dimensions for sampler compatibility.

The optional mask input is the neat one: mask out a region and only faces detected inside that mask are returned - "only the faces in the doorway, not the crowd."

Outputs

face - the cropped image(s) as a batch. crop_data - a CROP_DATA structure with the box and original context. That second output is the enabler: it feeds nodes that expect CROP_DATA to composite the result back into the original frame (the same convention used across the wider face-crop ecosystem), so your swapped/restored face can go home where it started.

Installing it

Pack standard: ComfyUI Manager → search comfyui-AGSoft → install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Art-xmaster/comfyui-AGSoft.git

The one real dependency is opencv-python (for image handling) - the pack's pyproject.toml lists it, but the plain requirements.txt doesn't, so if the node is missing after a bare clone, run pip install opencv-python in your ComfyUI environment and restart. RetinaFace weights are already in the folder; nothing to download.

Common issues

• Node doesn't appear after install - almost always missing cv2. Install opencv-python (above).
• Fewer faces than expected - raise number_of_faces or lower confidence_threshold; the default 0.7 drops small/partial faces on purpose.
• Crops too tight or too loose - that's scale_factor, not the detector. Bump it up if you're getting forehead-shaved crops.
• Face-swap identity drift - remember the crop feeds a face-focused pass (swap or restore) whose result goes back via crop_data; generative models rewriting faces is a pipeline design decision, not a node bug.

One honest note: RetinaFace is a 2019-era detector - great for clean frontal faces, weaker on heavy angles and tiny faces. For those cases, expect to lean on the confidence and scale knobs.