Have a graded image and the .cube file that was applied to it? This reverses the LUT and recovers the ungraded original. It runs entirely in your browser, and it tells you up front how much of the image can honestly be recovered.
Step 1
Step 2
A LUT is a lookup table. It takes a colour in and gives a colour out. To undo one, you take the graded pixel and search for the input colour that the LUT would turn into it. When that input is unique, the recovery is exact, down to the last of 256 levels.
The problem is that LUTs are not always one to one. A heavy contrast curve pushes many dark values to pure black. A greyscale conversion sends every hue of the same brightness to one grey. When several different colours arrive at the same output, the original cannot be worked out from the result. The information is gone, not hidden, and no amount of processing brings it back.
That is why the tool scores your LUT before you upload anything, and why the result reports what share of pixels it recovered confidently. Recovery tends to be all or nothing per pixel rather than uniformly blurry: a colour either comes back essentially exact, or it was never recoverable.
Mild creative looks, white balance shifts, gentle contrast curves, and most film emulation LUTs reverse cleanly, because they move colours around without collapsing them together. Log to Rec.709 conversions usually reverse well too, since they are designed to be invertible.
Aggressive contrast, crushed blacks, blown highlights, and anything monochrome do not. If the tool warns you that a LUT is barely reversible, believe it. You will get an image back, but it will be a plausible guess rather than your original.
The output is a reconstruction, not your original file. If the graded image was saved as a JPEG, it also carries compression artefacts that were not in the original and cannot be removed here. The most useful way to treat the result is as a clean starting point for re-grading, rather than as a recovered master. If you still have the original file, use that instead.