Inverted colours of generated images

[maxikubizek]

Hello,

i am working with my own dataset.
All pictures (train, test) have a white background.
However, during training the colours seem to invert (white background gets black, dark parts get bright, bright parts get dark,..) and sometimes change back to normal (check image: for 600,800,1000kimg its normal).
If the last generated good images are inverted (e.g. for 3000kimg), the generated defect images are inverted as well (check image).
Can anyone help how to keep the background white?

generated good images during stage 1:

test images:

train images:

generated inverted image: