半透明抠图 / Unmult Matting

Every neural matting model is guessing. Unmult doesn't guess - it does algebra. 半透明抠图 / Unmult Matting is the After Effects Unmult trick wrapped as a ComfyUI node: if your render sits on a known solid background, the composite equation C = αF + (1−α)B has exactly one solution for the foreground color F and the alpha α, and this node solves it. No model inference, no GPU, and it keeps every bit of the semi-transparency a segmentation network physically cannot represent.

That last point is the real reason you'd reach for it. Glow, fire, smoke, particles, lens flares, UI effects - the whole category of "bright thing on a pure-color background" - comes out with fractional alpha intact, in milliseconds, on a batch of video frames if you want. The README's own reconstruction test shows the math recovers both alpha and foreground color from a known composite with an average error of 0.0000 on black, white, and green-screen backgrounds. It's exact by construction, not by tuning.

Inputs and outputs

• image - the composite to unmult. Accepts batches, so sequence frames run through in one pass.
• bg_color - the background color as #RRGGBB. Common picks: #000000, #FFFFFF, #00FF00 (green screen), #FF00FF (magenta).
• 黑点 / 白点 (black point / white point) - two sliders. Below the black point, alpha is forced to zero (kills background noise residue); above the white point, alpha snaps to one (makes the subject solid). The defaults - 0 and 1 - are a pure pass-through.

Outputs: rgba_image (4-channel IMAGE, premultiplied composite ready for compositing) and alpha (MASK) if you want to drive something else with it.

The internals are straightforward: it measures how far each pixel deviates from the background color, normalizes per channel, takes the max as alpha, then solves foreground = bg + diff/alpha. The sliders are just a remap of that alpha before solving.

The trap: black background + dark solid subjects

This is the one that bites. Unmult on a black background is effectively treating brightness as opacity - which is exactly right for a glow and exactly wrong for a solid object. The README shows a solid subject whose true alpha is 1 everywhere: on a black background the bright parts reconstruct at 0.900 alpha but the dark parts (hair, dark clothes, shoes) fall to 0.080 - cut clean through. Same subject on green screen: 0.940 everywhere. So:

• Light effects, fire, smoke, particles, UI elements on pure color → this is the best tool in the pack.
• Green/magenta screens with high contrast against the subject → also great, even for solid subjects.
• Black background + solid character → use Lucida or FeyNobg instead. No amount of slider twiddling fixes a geometry problem.

If you're left with background fuzz after a good unmult, nudge the black point up a little rather than the white point down - that clears residue without hardening the edges that matter.

Installing

Part of RUI-Nodes:

ComfyUI Manager → search "RUI-Nodes" → install, restart
# or:
cd ComfyUI/custom_nodes
git clone https://github.com/rui40000/RUI-Nodes
cd RUI-Nodes && pip install -r requirements.txt

No weights to download - it's numpy math. The ecosystem's habit is to throw a 1GB model at every cutout problem, and this node is the useful counter-example: sometimes the right tool is a formula.