Just looked into this a bit. The proper way of fixing this would be to adapt the implementation in Camera.display_image_mobject -- basically, ImageMobjects are placed by specifying the coordinates of the upper left, upper right, and lower left corner (in ImageMobject.points). Manim is able to handle the case where the lower left corner is higher than the upper right corner (i.e., a flip along the horizontal axis), but curiously not when the upper right corner is left of the upper left corner (vertical flip).

I believe the problem mentioned by @behackl also occurs frequently in ThreeDScenes and leads to a rather comical effect. When an image is part of a ThreeDScene the flipping happens naturally during rotations causing the image to distort and flip wildly.

class ImageIn3D(ThreeDScene):
    def construct(self):
        image = ImageMobject(
            np.uint8(np.random.randint(256, size=(256, 256)))
        )
        image.rotate_about_origin(PI/4, axis=np.array([0., 1., 0.]), axes=[])
        image.shift(3*RIGHT)
        self.set_camera_orientation(phi=PI/3, theta=PI/4)
        self.add(Cube(fill_opacity=1.0))
        self.add(image)
        self.begin_ambient_camera_rotation(rate=PI/2)
        self.wait(4)
        self.stop_ambient_camera_rotation()

I believe this issue was fixed for the OpenGL renderer with #2534 as seen below. The distortion shown above and the flipping issue mentioned earlier seem to work now when OpenGL is used. The issues still exist with Cairo.