👁️ Computer Eyes
Dashboard 02 · Module 1

Alpha & Matte Lab

Every cut-out symbol you have ever approved lives or dies on a fourth number nobody looks at. Change the backdrop, harden the matte, break the premultiply — and watch the halo appear. This is the dashboard that explains every "why does it look dirty on the dark reel?" ticket you have ever received.

Symbol

Backdrop (the reel behind it)

Matte shaping

Push it up and soft edge pixels get forced to 0 or 255 — a 1-bit matte. This is what a hard "remove background" does. Watch the stair-steps.
Morphological erode (−) / dilate (+), 1 px per step. Spreading past the original edge reveals whatever RGB the file was hiding under α = 0 — usually junk, because most exporters do not clean it. That is not a bug in the lab; it is what is actually in the file.

Compositing maths

"Broken" bleeds the old background colour into every semi-transparent pixel — exactly what happens when a symbol is exported from a white-background comp without proper matting. Pick White studio or Green screen for a bright halo. Pick Black and you get the other classic: that is literally src×α, i.e. premultiplied data mistakenly tagged as straight — the dark fringe you see when an engine or exporter gets the alpha convention wrong.

Inspect

The one equation behind every composite

out = src × α + bg × (1 − α)
This is the over operator written for straight (unassociated) alpha — what PNG stores and what this lab uses.
If the file is premultiplied (associated), RGB is already scaled by α and the operator is
out = srcpm + bg × (1 − α) — no multiply. Applying the wrong convention is the bug.
For inspectable arithmetic this lab mixes encoded sRGB byte values; physically correct light mixing first linearises RGB.
Hover a pixel on the edge of the symbol to see the arithmetic run.
α = 1 → you see only the symbol. α = 0 → only the reel. Everything in between is a blend, and blending is where cut-outs go wrong: if the stored colour is contaminated, the blend prints the contamination.

Matte quality report

—
Semi-α % depends on padding and shape occupancy, so it is a poor portable score. Mean edge width (semi-α pixels ÷ perimeter pixels) is less padding-sensitive, but it is still measured in pixels and therefore changes with resolution. Compare like-for-like output sizes and inspect at the intended display size. 0% semi-α means either a hard cut or a fully opaque asset; only the former is a matte defect.
Set backdrop to White, look at the edge, then switch to Near-black. Now hit Broken. The symbol did not change — one number per pixel changed how it mixes. The halo you have been calling "bad AI cutout" is arithmetic: colour that should have been multiplied out of the edge was left in, and the dark reel exposes it. You can now specify the fix in one sentence instead of sending it back three times.