Color Image Fill Rm

Every now and then you get a mask or an overlay that's the right shape and the wrong color. A segmentation output that fills its subject in red instead of white. An annotation layer with boxes in a color your pipeline chokes on. A generated image with a background tone you want to swap for a cleaner one. Normally that's a trip to GIMP or a five-line PIL script, round-trip and back. ColorImageFillRm does it in the graph, in one node, on a whole batch - and it can also make a color fully transparent in the same pass.

It's from comfyui-missed-tool, the small utilities pack by ifmylove2011. The README's framing is honest about the sweet spot: "useful in certain scenes where mask images are overlaid with layers." Think of it as an in-graph chroma-key trimmer for whatever colors your intermediate steps introduced.

How it works

Pure torch, no PIL round-trip. The node looks at the shape of your batch:

• If the input already has an alpha channel, it keeps it. If not, it appends an opaque one on the fly.
• For every color you list in color_src, it finds matching pixels (within the threshold) and paints them with the matching color from color_replace.
• Then it takes rm_color, finds pixels matching that, and zeroes their alpha - that's the "Rm" in the name, removal to transparency.

The threshold is an Euclidean distance in normalized 0–1 RGB space, and the author's tooltip is refreshingly concrete: a default of 0.05 ≈ about ±13 levels per channel. Small number, small tolerance.

The inputs that matter

• images - the IMAGE batch you're cleaning up.
• color_src - the colors to find, hex with #, comma-separated (default #ffffff,#ff0000).
• color_replace - the swaps, same comma-separated hex format (default #00ffff,#0000ff). The count must match color_src, or the node throws a ValueError.
• rm_color - the color to make transparent (default #000000).
• threshold - match tolerance, default 0.05. If your target colors bleed or your image is soft, nudge it up; if you're getting unintended swaps, nudge it down.

The single output, image, is always RGBA after processing - even if your input was plain RGB, it comes out with an alpha channel.

Installing it

Grab it via ComfyUI Manager by searching comfyui-missed-tool, or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/ifmylove2011/comfyui-missing-tool
cd comfyui-missing-tool
pip install -r requirements.txt

Restart, and the node sits under the "missed-tool" menu section. Dependencies are just torch, numpy, Pillow - all already in a stock ComfyUI - and there are no model downloads to babysit. (The README's clone URL uses "comfyui-missed-tool"; the repo is "comfyui-missing-tool", and GitHub redirects either spelling.)

Where people get burned

• Counts must match. Two colors in color_src and three in color_replace? Hard error. Keep them paired one-to-one.
• The output has alpha. If the next node in your chain is expecting plain RGB, it'll silently drop the channel or behave like an RGBA tensor - wire a Composite or a dedicated RGBA consumer in if that matters.
• Threshold is small and normalized. 0.05 is roughly 13/255 per channel. Colors that look close to you in a thumbnail may be outside the tolerance; that's usually a good thing, but expect to tune it on your first real job.
• Replacement colors are opaque. Color swapping never touches alpha - only the explicit rm_color pass punches pixels out.

For mask cleanup and color-driven compositing it's a genuinely handy tool, and it being in-graph means the fix is reproducible the moment you save the workflow.