While training the model, are you converting the image in some range value or directly using the value from (0, 255)?