Censor with Mask

The node that obeys your mask

Censor with Mask is exactly what the name promises: give it an image and a mask, and it blurs, pixelates, or paints over the region the mask covers - nothing else. It's from the same pack as the detectors, but it does zero detecting on its own. The mask is on you, and that's the point. It slots into a pipeline where something upstream decides what to hide: typically Imgutils Generic Detector in "Censor Detection" mode finds the region, SAM Predictor refines it into a precise mask, a Mask Morphology node cleans that up, and this node does the hiding. The pack ships a censor example workflow wired exactly that way.

How it works

The three modes are unglamorous PIL/CV operations under the hood, which is fine - this is not a place that needs AI.

• blur - Gaussian blur with a blur_radius, composited back only inside the mask.
• pixelate - downscales the image by pixelate_size, then upscales with NEAREST. That's the classic mosaic.
• color - fills the mask region with a solid color_hex value at color_opacity.

The color mode is the one people sleep on. Crank color_opacity down to 0.4 and you get a tint instead of a slab, which looks far less jarring than a hard black rectangle when you're hiding a watermark or a stray hand.

Inputs and outputs that matter

Required: image, mask, and censor_mode (blur / pixelate / color). The three optional fields only matter for the mode you picked:

• blur_radius (default 5, 0.1–50) - blur mode.
• pixelate_size (default 10, 1–100) - bigger is chunkier blocks.
• color_hex and color_opacity (default #000000, 1.0) - color mode.

Output is a single censored_image. One minor gotcha: the blur radius is cast to an int internally, so 1.9 becomes 1, not a rounded 2. Doesn't matter much in practice.

Install

It's part of the LK-168/comfyui_imgutils pack. ComfyUI Manager → "Install Custom Nodes" → search "comfyui_imgutils", or:

cd ComfyUI/custom_nodes
git clone https://github.com/LK-168/comfyui_imgutils

Restart ComfyUI. The pack pulls torch, numpy, opencv-python, Pillow, segment_anything, and dghs-imgutils[gpu] - Manager handles those; manual installers may need pip install -r requirements.txt. This node itself is pure image math, but the detection and SAM nodes it's designed to chain with download models to $HF_HOME on first run, so set that env var before you start (the README is explicit about it).

Common issues

• Mask doesn't match the image. If your mask came from a different resolution or a cropped image, the censored region is in the wrong place. Resize the mask to the image's dimensions upstream.
• "Nothing changed." Check the mask isn't empty - run Mask Info from this same pack: 0% coverage means an all-black mask, and no censor mode will save that.
• Blurry blob around your subject. That's the mask being loose, not the node being broken. Tighten it with a morphology/edge operation before feeding it in.

Reach for this when you want selective hiding without rebuilding it yourself - it's one of those small nodes that makes the whole detection side of the pack feel finished.