Looks nice, but I think you might have some problems in the convolution step:
there’s 1k samples here, and you can see some very suspect pixels around the right leg. One of those (the left-most) popped white around 900 sample count, after having relatively common gradient with the neighbouring pixels.