The transparent checkerboard was hiding 5.73% of the frame

A developer comparing four background-removal paths on a challenging test image found that the largest AI model left 5.73% of the frame at alpha 1–8, an invisible residue on checkerboard that reappears on white. The non-AI path produced zero such pixels, while the three AI tiers differed mainly in where they placed uncertain regions: the fast tier discarded them, the pro tier smeared them over the subject, and the max tier spread them over the background.

A transparent checkerboard is a dishonest background. The same cutout looks spotless on it, and the moment you drop it on white, something from the original background fades back in. I ran into this while comparing three matting models, and my first guess was that I had written the compositing wrong. The test image was picked to be annoying. A hand holds a frosted glass bottle at an angle, the cap is mirror-finish metal, and behind it sits a pile of silver-white ornaments. Subject and background share a colour family, and the highlights and semi-transparent areas run into each other. There were four paths available: one non-AI "fast background removal", plus three AI tiers at roughly 42 MB, 219 MB and 447 MB. I wanted to know in what order you are supposed to choose between them. Ten minutes of staring at the four results side by side and I could not rank them. My assumption that they would line up neatly by file size just did not survive contact with the output. I don't own the matting algorithm — the codec and model-loading layer is my part of this — so instead of guessing from the picture, I read the channel. Here is the counter-intuitive one. In the heaviest tier's output, 5.73% of the frame sits at alpha 1–8 . At that opacity it is invisible on a checkerboard. Put it on solid white and the outline of those background ornaments comes back. The full distribution across the four paths: | alpha = 0 | 1–8 | 9–127 | 128–254 | 255 | | |---|---|---|---|---|---| | Fast removal non-AI | 59.44% | 0.000% | 1.77% | 1.56% | 37.22% | | Fast · ISNet INT8 | 75.66% | 0.72% | 0.45% | 2.04% | 21.13% | | Pro · BEN2 FP16 | 68.31% | 0.89% | 0.41% | 10.99% | 19.41% | | Max · BiRefNet HR-Matting | 63.13% | 5.73% | 1.23% | 5.84% | 24.07% | One image, desktop Chromium. These numbers describe this photo and nothing else. Look at the second column: the non-AI path is a flat 0.000%, not a single pixel, because it isn't a model — it deletes background colour that is both similar and connected, starting from the canvas edge, so a pixel is either in or out with no middle state. All three models trail an alpha 1–8 tail, and the largest model's tail is nearly eight times the other two. Pixels above alpha 10: 23.6% for the fast tier, 30.8% for pro, 31.1% for max. The fast tier's non-zero region is boxed inside x 270–790 and y 129–1341 — the left two hundred columns and the bottom ninety rows are completely empty. That is not a badly finished edge. The model decided the subject of this photo is the bottle and the hand is background. Sampling the hand region gives a mean alpha of 102, with a few fingernail outlines still hanging on. If what you wanted was a bottle on its own, that is exactly right. If "held in a hand" was the point of the shot, the model has redefined the subject rather than done a rough job on it. The pro tier keeps the hand and pays for it in the fourth column: 10.99% of pixels land in 128–254, and pulling them out shows almost all are 249–254, a very light soft alpha over the whole subject that looks opaque but isn't 255. The max tier gives the subject the highest share of hard 255 of the three, then parks its uncertainty as that faint 5.73% residue spread back across the background. The three tiers don't differ in edge sharpness. They differ in where they put the parts they aren't sure about: the fast tier throws them away, the pro tier smears them over the subject, the max tier smears them over the background. Which means the first question isn't "which model is strongest", it's "what kind of mistake can this image afford". Uniform-colour backgrounds — packshots, flat-colour assets — should try the non-AI path first and skip the download entirely. Its failure mode is easy to spot: any disconnected chunk of background survives whole and you see it immediately. Simple subject, single object, clean background: 42 MB is enough, and it is the only tier that doesn't make a low-spec machine wait. Hands, hair, semi-transparent material — go up, but add a step after the biggest tier, because that residue only shows on a solid background and you'll ship it otherwise. One thing to fold into the choice: the models come down from the site and go into browser cache, inference runs locally, and the first time you pick the 447 MB tier you wait for all 447 MB. That cost is real even though it's a one-off. Don't max the slider out of habit. The tool I ran this on is ImgIng imging.ai .