Remove a background with a model or with a colour key that has a real alpha ramp, scored against a corpus whose right answer was chosen rather than found.
Initializing in your browser…
Remove EXIF, GPS, XMP and IPTC by rewriting the container rather than re-encoding the picture, so the pixels are byte-for-byte identical and the file gets smaller. Selective removal, batch mode and a ZIP.
Turn an image into a data URL whose payload is decoded again and compared byte for byte with the file, with the media type taken from the file's own signature rather than its extension and the size cost of embedding stated in full.
Compare two images with a perceptual engine (pixelmatch) that ignores anti-aliasing, or an exact byte engine. Groups changes into regions with bounding boxes, reports SSIM, and handles size mismatches with an explicit policy.
A known foreground composited onto a known background at a known alpha, so the right answer for every pixel is the number the compositing used.
Input
480 x 360 · 113,442 background pixels, 54,995 foreground, 4,363 partly both
What came back, scored against that alpha
threshold ramp model background removed 113,442 113,442 113,343 subject kept 54,995 54,995 48,040 soft edge recovered 0 537 990 worst pixel, of 255 211 95 255 seconds 0.1 0.1 12.4 on a graduated background background removed 67,034 113,442 113,243 mean error, of 255 69.93 5.30 12.31
The colour key was already perfect at the two things it is asked to do on a flat background, and that is recorded rather than claimed as a fix. What a hard threshold cannot do at all is a soft edge: a pixel that is part background and part subject sits at a middle distance and a threshold can only call it one or the other, so all 4,363 of them kept the background colour. The ramp is now fitted to the picture, and on a background whose corners differ the old four-corner average left 46,408 background pixels behind.
Remove a background with a model, or with a colour key that has a real alpha ramp fitted to the picture. The page reports how many pixels it removed, how many it kept and how many edge pixels came back partly transparent, and how long the model took.
Cut a product off its backdrop for a listing, with the fringe removed rather than left on the edge.
Let the model find the subject in a photograph where no single background colour exists.
Key a chroma green frame with a ramp, comparing colour alone so shadows survive.
Choose Color Based and press Remove Background. The corners give the backdrop colour, Otsu fits the ramp, and the panel reports how many pixels were removed, how many kept, and how many came back partly transparent.
Leave the mode on AI. The model downloads once and takes about twelve seconds, and the page shows the elapsed time when it finishes.
The page says the four corners differ by up to 40 of 255 and that their average may appear nowhere in the picture. Pick the colour you mean with the eyedropper, or let the fitted ramp handle it: the graduated fixture now comes back with every background pixel removed.
A background remover is one of the few image tools whose right answer can be CONSTRUCTED rather than found. Composite a known foreground onto a known background at a known alpha, save the alpha map beside the picture, and every pixel of the result has an answer that is arithmetic: the number the compositing used. That is the corpus this page is scored against, and it settled three things.
**The colour key was already perfect at the two things it is asked to do.** On a foreground composited onto flat chroma green it removed 113,442 of 113,442 background pixels and kept 54,995 of 54,995 subject pixels, and the same on flat white. Recording that is worth as much as a fault would have been, because it moved the pass to the two places the key really failed.
**A hard threshold cannot key a soft edge, by construction.** The alpha was set to 0 or left at 255 and nothing in between, so every pixel that was PART background in the original kept the whole background colour: of the 4,363 partly transparent pixels in the fixture it recovered none, and the worst pixel was 211 levels of 255 away from the truth. That is what a green fringe is. The alpha is a RAMP now, and the ramp is fitted to the picture rather than guessed: the two clusters are separated with Otsu’s threshold, the same one this project’s OCR preparation uses to binarise a page, and the ramp runs from the background’s own noise floor to the distance of the NEAREST part of the subject. Compositing is linear, so a pixel half way between the two in colour really is half transparent, and stopping at the nearest part of the subject is what keeps the subject itself solid. Measured after: 537 edge pixels recovered where there were none, the worst pixel down from 211 to 95, and the subject still 54,995 of 54,995.
**A background colour has to be found, not assumed.** The old mode averaged the four corners, which on a background whose corners are four different colours gives a colour that appears nowhere in the picture: it removed 67,034 of 113,442 background pixels and left the rest, at a mean error of 69.93 of 255. The corners are still sampled, but their DISAGREEMENT is measured and printed, and with the ramp fitted the same picture now comes back with all 113,442 removed at a mean error of 5.30.
Spill is suppressed by name rather than left in: an edge pixel is a mixture, so it carries the background’s colour, and removing that contribution in proportion to how transparent the pixel is takes the fringe with it. There is a choice of what to compare, too: colour and brightness together, which is what most tools mean, or colour alone, which ignores how bright a pixel is and so keeps a shadow on a green screen rather than cutting a hole in the subject.
The model is the other mode, and it is TIMED, because a background remover downloads tens of megabytes and that is exactly where a progress bar stops moving and nothing says why. Measured here it returns in about 12 seconds including the download, removes over 99 percent of the background whatever the background is, and produces a soft edge of its own: 990 of the 4,363 edge pixels came back partly transparent. On this synthetic corpus of flat colours it keeps less of the subject than the colour key does (48,040 of 54,995), which is what you would expect of a model trained on photographs of people and objects rather than on discs, and it is stated rather than hidden. Use the model on a photograph and the colour key on a flat or graduated backdrop.
The cut-out can then be put on a new background: transparent, a solid colour, a gradient, a blurred copy of the original, or another picture. Everything runs in the browser and nothing is uploaded.
The colour key for a flat or graduated backdrop: it is a tenth of a second and, measured against a corpus whose alpha was chosen in advance, it removed every background pixel and kept every subject pixel. The model for a photograph, where no single background colour exists.
Because it set the alpha to 0 or left it at 255 and nothing in between. A pixel that is part background and part subject sits at a middle distance from the background colour, and a threshold can only call it one or the other, so it kept the whole background colour. On the test fixture that was all 4,363 partly transparent pixels.
It separates the background and the subject with Otsu’s threshold, then runs the ramp from the background’s own noise floor up to the distance of the nearest part of the subject. Compositing is linear, so that recovers the fraction directly, and stopping at the NEAREST part is what keeps the subject solid.
About twelve seconds here, most of it downloading, and the page shows the number when it is done. A model that downloads tens of megabytes is exactly where a progress bar stops moving and nothing says why, so it is timed.
No. On a corpus of flat colours it kept 48,040 of 54,995 subject pixels against the colour key’s 54,995, because it is trained on photographs rather than on flat shapes. It is better at a picture with no single background colour, and it produces a soft edge of its own.
No. The model runs in your browser and the colour key never leaves the page.
Images are decoded, edited, and exported entirely inside this browser tab. No originals, exports, or metadata are uploaded.