🐳自动打码

If you've ever had to blur out faces in fifty generated images by hand, you know exactly why this node exists. 🐳自动打码 (AutoCensorWithOpenPose) runs pose detection, finds a person's face, chest, and groin regions, and applies a Gaussian blur to each - automatically, per image, in a batch. It's the "make these safe to upload" node, and it's the reason this pack got its name in some workflow-sharing circles. No API, no cloud, no key: it all runs locally.

The mechanism is MediaPipe under the hood, wearing an OpenPose costume. The node runs MediaPipe's pose landmarker to get body keypoints, then maps them into the OpenPose BODY_25 format (25 joints: nose, shoulders, hips, knees, eyes, ears…). From those landmarks it computes where the face box is (between the eyes and nose), where the chest is (between shoulders and mid-hip), and where the groin is (around the mid-hip landmark), then blurs those rectangles. The README notes that mediapipe is an optional dependency with a fallback, but honestly - the fallback path is thin and the pose-based regions are the whole point. If you want this to work well, let pip install -r requirements.txt pull in mediapipe.

The inputs

• censor_face / censor_chest / censor_groin - three toggles, default off/on/on. Face is off by default, which is a deliberate choice: plenty of workflows want the body censored but the face left alone (or vice versa).
• blur_strength - 5 to 50, default 20. The bigger the blur radius, the more the region dissolves. 20 is a reasonable middle; you'll want higher if the source is high-res.
• censor_size_multiplier - 0.8 to 2.0, default 1.2. Scales each censor region up or down. Crank it if the blur keeps clipping body parts, or drop it if the boxes swallow too much.

Input is image, output is 打码后的图像 (the blurred IMAGE). Single input, single output, no mask output - so if you want the mask itself for something else, you're out of luck here.

Installing and using it

Standard Pond Nodes install - Manager search comfy_Pond_Nodes, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Pondowner857/comfy_Pond_Nodes
cd comfy_Pond_Nodes && pip install -r requirements.txt

Two things worth knowing before you trust it on a big batch. First, MediaPipe is not an identity model - it finds a face, it doesn't tell you whose face. If you need recognition, that's a different stack (the KB's detailing docs make this exact point). Second, it's a pose-driven box, not a segmentation: the blur regions are rectangles around landmarks, so on unusual poses (side-on, seated, cropped bodies) the boxes drift. Run it on a couple of test images first, tune censor_size_multiplier, then let it loose on the folder. Also keep in mind this pack is a big Chinese-authored kitchen sink - if the node is missing at search time, make sure Manager pulled the latest version, since the pack updates in large feature drops.