You could avoid the clipping by normalizing the images first:
images -= images.min()
images /= images.max()
I just created the images using gaussian noise (torch.randn).
You should use your images tensor instead or do you see random noise there too?