Stamp Censor

Stamp Censor is the whole reason the pack exists. Where a mosaic or blur node pixelates a region into a gray smear, this one scatters a stamp shape over the mask until it's covered - hearts, scribble bars, stars, crosses, or your own PNG. Same mask pipeline, completely different result. It's the censorship you run when you want the result to look deliberate rather than degraded.

It fits into your workflow exactly where a blur/mosaic node would. You already have a detection chain producing a white-on-black mask (YOLO via Impact Pack, SAM3, or a Grow Mask node) - you just feed that mask here instead of into a blur. The pack is emphatic about this: it does not detect anything and it does not grow masks. Wire those nodes yourself. What you get is the prettiest possible way to cover what they found.

How it works

Under the hood it treats each connected white region of the mask as a separate job. For each region it computes the region's long-axis heading, then places stamps at random sizes, rotating them to follow that heading (that's auto_rotate) and packing them with oriented-box collision until target_coverage of the region's mask is covered. Tiny islands - the SAM speckles that would get one lonely stamp - are skipped automatically. Stamps are allowed to overflow the mask edges, which is what makes the coverage look natural instead of like a crop. If the connected stamp is a Stamp Load, the node preview on the canvas updates live, so you're not queueing to see if the layout works.

The inputs that matter

image, mask, and stamp are the three wires: the source image, the white-on-black mask, and the STAMP output from Stamp Load or Stamp Custom Load. The rest are the knobs, and only a few are worth touching on a first pass:

• size_ratio (default 0.28) - stamp size as a fraction of the region's shorter side. 0.28 is a good "clearly a stamp, not a smear" size; push toward 0.5–1.0 and the region gets a few big stamps instead of many small ones.
• target_coverage (default 0.8) - stop placing stamps when this fraction of the mask is covered. Lower it and you get sparse, obviously-intentional stamping; higher and it approaches a solid block.
• spacing_factor (default 0.3) - how much stamps are allowed to overlap. 0 is full overlap, 1.0 is "oriented boxes just touch."
• size_jitter (default 0.15) and angle_jitter (default 0) - randomness so the layout doesn't look like a spreadsheet.
• auto_rotate (on by default) - align each region's stamps to its long axis, then add the load angle.
• uniform_pack (off) - the "Even Pack" mode: instead of random samples, slender regions get a regular scan along the long axis and squat regions a brick-staggered layout. Turn this on when the random scatter reads as messy and you want the neat version.
• seed (optional) - leave empty and the layout randomizes every queue; connect one to keep a layout reproducible.

Output is a single image (IMAGE), ready for a Preview Image or Save node.

Install

Search 4A Stamp Censor in ComfyUI-Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/tsukino4a/ComfyUI-4A-StampCensor.git ComfyUI-4A-StampCensor

Restart ComfyUI after. The pleasant surprise: no extra pip packages, no model downloads. Pillow and NumPy come from ComfyUI, and the stamps are plain PNGs. This is about the lightest install in the ComfyUI ecosystem.

Common gotchas

The #1 beginner stumble is mask polarity - white areas are stamped, black is left alone, and if you feed it a mask from a node that outputs black-on-white, you get... nothing, or the whole frame. Check the mask output before blaming this node. Second: if nothing appears, the stamp input isn't connected (or is connected to something that isn't a STAMP). Third: it does not detect NSFW or grow masks, so if your mask is a tight bounding box you'll get a rectangle of stamps - that's on the detection chain, not here. Sample workflows ship in example_workflows/ (SAM3 and Impact Pack YOLO versions) if you want a working reference.

If you're making something to share or sell, this is the censor node you'll reach for - it's deterministic, cheap, and makes the covered region look chosen rather than censored.