Context :

Developing tests for a library generating complex images with canvas (see example image). An exact image match (by the pixel) is needed. I am trying to automate the regression tests on the library.
sample generated image

Question : should I try further or is there a documented imprecision of the “difference” blend mode ?

Process :

STAGE A. STORE VALID DATA

1. generate a canvas with the code to test, using a set of generation parameters.
2. store resulting image data as a png in the filesystem.

STAGE B. GENERATE NEW DATA

1. code evolves, there is a need to retest.
2. generate again, with same parameters as in A.1. above.

STAGE C. COMPARE A and B DATA

1. create a new canvas (or clone the generated one.)
2. put the last generation data in the new canvas.
3. set the mix-blend-mode of the new canvas to “difference”.
4. load a new image (async/await) with the reference generation data (from step 2.)
5. put the reference generation image (from step 2) in the new canvas.
6. check the resulting image data (each pixel data should be 0 for RGB or 255 for alpha) :

function pixelDataIsEmpty(value, index){
    return (value !== 0) && (!((value == 255) && ((index + 1) % 4 == 0)));
}

Problem :

Some ghost differences are generated, despite imageData for A & B look identical by the pixel when echoed in the console .
Very oddly, this happens only because of step C.1-2. When applied directly the new canvas generated in B.2., the exact same function works, the difference blend behaves nicely.